This is related to U.S. Application entitled “Transaction Grouping and Rescheduling in a Database System,” Ser. No. 10/121,191, filed concurrently herewith by Gang Luo, Curt J. Ellmann, Jeffrey F. Naughton, and Michael W. Watzke.
A database is a collection of stored data that is logically related and that is accessible by one or more users. A popular type of database is the relational database management system (RDBMS), which includes relational tables made up of rows and columns. Each row represents an occurrence of an entity defined by a table, with an entity being a person, place, or thing about which the table contains information. To extract data from, or to update, a relational table, queries according to a standard database query language (e.g., Structured Query Language or SQL) are used. A table (also referred to as a relation) is made up of multiple rows (also referred to as tuples). Each row (or tuple) includes multiple columns (or attributes).
Various other data structures are also typically associated with relations in a relational database system. For example, a view is a derived relation formed by performing a function on one or more base relations. Rather than storing the view, the function is typically recomputed each time the view is referenced. This type of view is referred to as an “ordinary view.”
Unlike an ordinary view, a materialized view is a pre-computed, stored query result that can be used for some queries instead of reconstructing the results directly from the base relations. As with the ordinary view, a function is performed on the base relations to derive the materialized view. However, because the materialized view is stored, fast access to the data is possible without recomputing the view.
After the materialized view is created, subsequent queries are able to use the materialized view, where appropriate, to increase query processing speed. Materialized views can be used to assemble data that come from many different relations. One type of view is the join view, which stores join results of multiple base relations.
A materialized view is updated when the underlying base relations are modified. As the base relations are changed through insertion of new tuples, deletion of tuples, or updates to existing tuples, the corresponding rows in the materialized view are changed to avoid becoming stale. This is known as materialized view maintenance.
Relational database systems can be used for data warehouse applications. A data warehouse collects information from several source databases. The collected information is integrated into a single database to be queried by the data warehouse clients.
Traditionally, data warehouses have been archival stores used for analysis of historical data. More recently, however, there has been a growing trend to use a data warehouse operationally (referred to as a “operational data warehouse” or “real-time data warehouse”), which involves making relatively real-time decisions about data stored in the data warehouse.
Traditional techniques of maintaining views are usually inadequate (in terms of processing speed) for operational data warehouses due to the real-time update requirements. Furthermore, materialized view maintenance in an operational data warehouse requires transactional consistency. If transactional consistency is enforced by traditional concurrency control mechanisms (including locking mechanisms), the ability of the database system to perform concurrent transactions may be reduced. This hurts performance in a database system, especially in a parallel database system having multiple processing modules.
A further issue involved in a database system is the occurrence of deadlock among different transactions. For transactional consistency, each transaction in a database typically places some type of lock on relations and views that are involved in the transaction. In some scenarios, the conflicting locks for multiple transactions are placed in the relations in such an order that none of the multiple transactions can proceed further—a deadlock condition. A deadlock among transactions reduces the ability of transactions to successfully complete in a database system, and as a result, system performance suffers.
In general, a mechanism for rescheduling or reordering transactions is provided to reduce the likelihood of lock conflicts and/or deadlocks between transactions in a database system. For example, a method of scheduling transactions in a database system comprises receiving pending transactions, storing transactions in a queue for execution, and identifying one of the received pending transactions that does not conflict with any of the transactions in the queue. The identified transaction is moved into the queue.
Other or alternative features will become apparent from the following description, from the drawings, and from the claims.
In the following description, numerous details are set forth to provide an understanding of the present invention. However, it will be understood by those skilled in the art that the present invention may be practiced without these details and that numerous variations or modifications from the described embodiments are possible.
I. Introduction
In accordance with some embodiments of the invention, a transaction rescheduling mechanism is used to reduce the number of occurrences of lock conflicts and/or deadlocks between or among transactions in a relational database system. In many database applications, materialized views (such as join views) are provided to improve database query processing, such as join processing. A common form of materialized view is the join view, which is a view that stores and maintains the result of a join of multiple base relations (also referred to as “base tables”). Although reference is made to “materialized views” in some described embodiments, the more general term “view” is intended to cover both ordinary views and materialized views. An ordinary view is a view that is not persistently stored in database storage, but rather is recomputed for each newly received query that involves the view.
The introduction of materialized views into a database system brings with it an increase in the number of lock conflicts that may arise due to write transactions. The conflicts are due to the dependencies introduced among base relations by each materialized view. When a base relation is updated, the corresponding materialized view must also be updated, and that update requires a read of the other base relations used by the materialized view.
The following provides an example query for creating a join view (JV) on two base relations (A, B):
The join view JV includes tuples (also referred to as “rows”) of base relations A and B where the attributes (also referred to as “columns”) A.c and B.d are equal. The join view JV is partitioned on the attribute A.e. In other examples, a join view can be stored for a join of more than two base relations. In the following description, the terms “table” and “relation” are used interchangeably. Also, a table or relation has rows (or tuples) and columns (or attributes). The terms “rows” and “tuple” are used interchangeably, and the terms “column” and “attribute” are used interchangeably.
The join relations A and B and the join view JV created as a result of the query is shown in
There are various different types of locks that can be placed on data stored in relational tables to restrict access to or the ability to modify the data. Table-level locks are placed on an entire table or relation. Table-level locks include a table-level shared (S) lock and a table-level exclusive (X) lock. Generally, once placed on a table, a table-level S lock blocks a subsequent transaction that attempts to write to any part of the table. A table-level X lock placed on a table is more restrictive, as it blocks any subsequent transaction that attempts to read from or write to any part of the table.
While a table-level locking mechanism locks an entire table, a value locking mechanism locks only a portion of the table. The value locking mechanism specifies a value (or values) of an attribute(s) in a table for which locks are to be placed. Such an attribute, or attributes, is referred to as a value locking attribute, or value locking attributes. A value locking mechanism usually locks only one row or a few rows.
Value locks include a shared (S) value lock and an exclusive (X) value lock. To place an X value lock on an attribute value of a base relation R, a table-level intention exclusive (IX) lock is first placed on R. Similarly, to place an S value lock on an attribute value of the base relation R, a table-level intention shared (IS) lock is first placed on R.
A table-level IX lock is placed on a relation to prevent any subsequent table-level X or S locks on the same relation. The IX lock is a mechanism for the database system to efficiently determine whether a subsequent table-level X or S lock can be placed, without having to find out if there is a conflicting value lock on the relation. For example, suppose the value locking attribute of a relation R is attribute R.a. There can potentially be multiple value locks placed on multiple values of R.a. Thus, a first X value lock can be placed on row(s) of the relation R with R.a=5, a second X value lock can be placed on row(s) of relation R with R.a=2, and so forth. If a subsequent transaction attempts to read the entire relation R, the subsequent transaction will need to acquire a table-level S lock. One way to determine if the table-level lock S lock can be granted is by finding each R.a value for which there is an X value lock. However, this is inefficient. Instead, according to some embodiments, a table-level IX lock is placed on the relation R if there is at least one X value lock on the base relation R. Thus, to determine if the table-level S lock can be placed on the relation R, the database system can quickly detect that there is already an IX lock placed on the relation R, which blocks acquisition of the table-level S lock. Note that only one IX lock is needed for multiple X value locks of a relation. The IX lock also blocks any subsequent table-level X lock.
Similarly, a table-level IS lock placed on a relation R for an S value lock blocks any subsequent table-level X lock.
If a transaction T attempts to update base relation R, the transaction has two choices: (1) T can place a table-level X lock on base relation R; or (2) T can place an IX lock on R and an X value lock on some value locking attribute of R. Similarly, if transaction T attempts to read base relation R, transaction T also has two choices: (1) T can place a table-level S lock on R; or (2) T can place an IS lock on R and an S value lock on some value locking attribute of R.
If materialized views are also stored in the database system, a locking mechanism is also provided for the materialized view. In one embodiment, a “Y-lock” locking mechanism is provided for the materialized view. In another embodiment, a “no-lock” locking mechanism is provided for the materialized view. In yet another embodiment, a “name-lock” locking mechanism is provided for the materialized view.
These locking mechanisms are designed to reduce the likelihood of lock conflicts between transactions that operate on a materialized view. Although such locking mechanisms on materialized views decrease the likelihood of lock conflicts due to the presence of materialized views, further reduction of lock conflicts is still desired by using a transaction rescheduling mechanism according to some embodiments. The transaction rescheduling mechanism according to some embodiments reorders pending transactions so that locks conflicts due to the presence of materialized views can be reduced. Additionally, query grouping can also be optionally used to group queries for the purpose of eliminating duplicate locks so that the number of locks placed on base relations and views are reduced.
In the following description, Section II describes an example database system architecture, and Sections III–VIII describe various locking mechanisms. Sections IX–X describe transaction rescheduling mechanisms according to several embodiments.
II. Example Database System Arrangement
According to one embodiment, each data server module 12 includes a lock manager 15 to provide a locking mechanism according to some embodiments of the invention. Thus, the lock manager 15 is responsible for placing locks (e.g. table-level locks or value locks) on base relations and join views. As shown, the locking mechanism is distributed across plural data server modules 12. Alternatively, a centralized lock manager is employed.
In one embodiment, the requesting client that sends commands to the data server modules 12 include one or more parsing engines 16. The parsing engine(s) 16 receive requests from users or applications, which are in the form of queries according to a standard database query language (such as a Structured Query Language or SQL, as provided by the American National Standards Institute or ANSI). In other embodiments, other types of database query languages can be used.
Each parsing engine 16 includes a parser 18 and an optimizer 20. The parser 18 checks a received request for proper syntax and semantically evaluates the request. The optimizer 20 develops an execution plan for the received request. In some embodiments, the optimizer 20 uses a cost-based mechanism to select a least expensive (in terms of system resources utilized) execution plan for the query.
The execution plan includes a series of “steps” that are communicated by a scheduler 21 in a determined sequence to one or more of the data server modules 12 over a communications network 22. If the execution plan can be executed by one data server module 12, then the parsing engine 16 sends the one or more steps of the execution plan to the one data server module 12. However, if plural data server modules 12 are involved in the execution plan, the parsing engine 16 sends the step(s) to the plural data server modules 12. The sequence in which steps of an execution plan are executed is controlled by the scheduler 21.
In the example shown in
Although the storage modules 14 are shown as separate modules, they can be part of the same storage subsystem. Alternatively, the storage modules 14 can be separate storage devices. In another embodiment, instead of plural data server modules 12, the database system 10 can include only one data server module 12. The locking mechanism according to the various embodiments can be applied to a single-data server module database system rather than the multiple-data server module database system shown in
III. Y-Lock Locking Mechanism
In accordance with one embodiment of the invention, the locking mechanism provided for the join view is a “Y-lock” locking mechanism. A Y lock has similar functionality as an X lock, with the major difference being that the Y lock does not conflict with itself. In other words, the Y lock is a modified type of exclusive lock placed on a join view that allows subsequent Y locks to be placed on the same join view even though a first Y lock is still active on the join view. In contrast, a table-level X lock placed on a join view blocks a subsequent X lock on the same join view. By enabling multiple Y locks to be concurrently placed on the same join view, throughput of transactions in the database system 10 can be greatly increased, as the possibility of lock conflicts on the join views among different transactions is reduced. A Y lock is also referred to as a “modified-exclusive” lock.
Three types of table-level locks are allowed on a join view (JV): Y locks, S locks, and X locks. The conditions under which such table-level locks are allowed are shown by the logic of
Upon receiving steps associated with a transaction, the data server module 12 first determines (at 102) if the transaction specifies both a read and write of the join view JV. As used here, writing a view is also referred to as updating or modifying the view. If the transaction specifies both a read and write of JV, then the data server module 12 determines (at 104) if an S lock, X lock, or Y lock is currently active on the join view JV. If none of an S lock, X lock, or Y lock is active on the join view JV, the data server module 12 places (at 106) an X lock on the join view JV. Note that the X lock required for the received transaction conflicts with any of an X lock, S lock, or Y lock, so that the presence of any such table-level lock on JV blocks the required X lock for the received transaction.
If the data server module 12 determines (at 104) that any of the S lock, X lock, or Y lock is active on JV, the data server module 12 performs (at 108) conflicting lock processing. For example, the data server module 12 can wait until the current conflicting lock has been released. Different embodiments perform different tasks in response to detecting that a received transaction cannot proceed because of conflicting locks placed on the join view JV.
If the transaction is determined not to involve both a read and write of the join view JV (at 102), the data server module 12 checks (at 110) if the transaction involves a read (but not a write) of the join view. If so, the data server module 12 checks (at 112) if there is either a Y lock or an X lock on JV. If neither a Y lock nor X lock is currently active on JV, the data server module 12 places (at 114) an S lock on the join view JV. Note that the required S lock conflicts with either a Y lock or X lock, so that the presence of either the Y lock or X lock blocks acquisition of the S lock for the received transaction. If either a Y lock or X lock is presently active on JV, the data server module 12 processes (at 108) the conflicting lock condition.
If the transaction does not involve a read of the join view JV, the data server module checks (at 116) if the transaction involves a write (but not a read) of the join view JV. If so, the data server module 12 checks (at 118) if there is either an S lock or an X lock on JV. If not, then the data server module 12 places (at 120) a Y lock on the join view JV. However, if either an S lock or an X lock is presently active on JV, the data server module 12 processes (at 108) the conflicting lock condition.
The discussion above refers to placing a table-level Y lock on a join view. The Y-lock locking mechanism for join views is extendable to also allow value locks on join views. Consider a join view JV defined on base relations R1, R2, . . . , and Rn. For a fixed i (1≦i≦n), suppose that Ri.ai is an attribute of base relation Ri that also appears in JV. This is the case for the example join view JV of
The IY lock is similar to the traditional IX lock except that it is compatible with a Y lock or another IY lock. As with the IX lock, the table-level IY lock is placed on the join view JV in conjunction with an X value lock of JV to indicate to subsequent transactions that table-level X or S locks on JV are blocked (however, a table-level Y or IY lock on JV is still possible in the presence of the IY lock with X value lock). Also, a subsequent IS lock with an S value lock on JV is allowed in the presence of an IY lock with X value lock on JV. Note, however, that an X value lock on JV is incompatible with an S value lock placed on the same attribute value of JV.
For a transaction that performs both a (table-level) read and (value) write of the join view JV, both an S lock and X value lock are needed on the join view JV. In this case, a table-level SIY lock (which is equivalent to an S lock and an IY lock) is placed on JV. The SIY lock is similar to the traditional SIX lock. One can think that IX=IS+IY. An SIX lock is equivalent to an S lock and an IX lock (for an X value lock). The Sly lock is only compatible with the IS lock.
Note that SIX=S+IX=S+(IS+IY)=(S+IS)+IY=S+IY=SIY. Thus, the SIX lock is the same as the SIY lock.
If transaction T both updates JV (without specifying the Ri.ai value(s)), which is a table-write, and reads JV (specifying the Ri.ai value(s)), which is a value-read, then transaction T requires both a Y lock and S value lock(s) on JV. In this case, a table-level YIS lock is played on JV (which is equivalent to a Y lock and an IS lock). The YIS lock (Y+IS) is similar to the SIX lock and is only compatible with the IY lock.
The compatibilities of the different locks are listed in Table 1.
According to Table 1, a Y lock on JV is compatible with another Y lock or an IY lock on JV. However, the Y lock is incompatible with a table-level S lock, X lock, IS lock, IX lock, SIY lock, or YIS lock. Note that a table-level X lock is incompatible with any lock. An IY lock on JV is compatible with a table-level Y lock, IS lock, IY lock, IX lock, or YIS lock. However, an IY lock is incompatible with a table-level S lock, X lock, and SIY lock. An IX lock is compatible with an IS, IY, or IX lock, but not with any other locks. An SIY lock (S+IY lock) is compatible with an IS lock, but not with any other locks. A YIS lock (Y+IS lock) is compatible with an IY lock, but not with any other locks.
To show that the Y-lock locking mechanism keeps the isolation property (serializability) of transactions, the following assertions are proven for a join view JV defined on base relations R1,R2, . . . , and Rn:
The four assertions are first proven for the simple case where JV=σ(R1. . .
Ri . . .
. . .
Rn), where σ denotes a selection operator. The assertions are also proven (further below) for the general case where JV=π(σ(R1
. . .
Ri
. . .
Rn)), where π is a projection operator. Selection refers to selecting rows of base relations to place into JV based on the rows satisfying the join condition. Projection refers to projecting only those attributes that are in the select list of the join query into the join view JV. Less than all attributes of each base relation can be projected into the join view TV.
It is assumed that join view JV allows duplicate tuples. If no duplicate tuples are allowed in JV, assume that each tuple in JV has a dupent (or some other named parameter) attribute recording the number of copies of that tuple. The following assumptions are made (the other more complex cases can be proven in a similar way): (1) any transaction T updates at most one base relation of JV; and (2) if transaction T tries to update base relation Ri(1≦i≦n), it places a Y lock on JV and an S lock on each Rj (1≦j≦n, j≠i). For example, if a join view JV is defined on base relations A, B, and C, an update of A causes an X lock or IX lock to be placed on A, and a Y lock to be placed on JV. In addition, an S lock is placed on each of the other base relations B and C.
If transaction T writes join view JV, T places a table-level Y lock on JV until T completes. If transaction T both reads and writes join view JV, T places a table-level X lock on JV until T completes. Thus transaction T's writes to join view JV are not read by any other transaction T′ until transaction T completes, since T′ requires a table-level S lock on JV (which would conflict with the Y lock or X lock on JV for transaction T). This proves the first part of Assertion 1.
Additionally, if transaction T writes join view JV, there are two possible cases:
Suppose that transactions T and T′ update ΔRi and ΔRi′ of base relation Ri, respectively. ΔRi refers to the changed portion of Ri made by transaction T, and ΔRi′ refers to the changed portion of Ri made by transaction T′. There are three possible scenarios:
Thus transaction T's writes to join view JV are not written by any other transaction T′ until transaction T completes (this proves the second part of Assertion 1).
Based on a similar reasoning to the proof of Assertion 1, transaction T does not overwrite dirty data of other transactions in join view JV (this proves Assertion 2).
Suppose that transaction T reads join view JV by requiring a table-level S lock on JV. If some other transaction T′ is writing any data in join view JV, T′ will place a table-level Y lock (or X lock) on JV until T′ completes. The required table-level S lock on JV for transaction T will be blocked by the table-level Y or X lock on JV for transaction T′. Thus transaction T does not read dirty data from transaction T′ in join view JV (this proves Assertion 3). The case that transaction T tries to both read and update the join view JV can be proved similarly, since the required X lock will be blocked by the Y or X lock for transaction T′.
If transaction T reads join view JV, T will place a table-level S lock on JV until T completes. If transaction T both reads and writes join view JV, T will place a table-level X lock on JV until T completes. Thus no other transaction T′ can write any data in JV until transaction T completes, since T′ requires a table-level Y lock (or X lock) on JV. This proves Assertion 4.
The proof for the general case where JV=π(σ(R1. . .
Ri
. . .
Rn)) is discussed here. When projection is applied, less than all the attributes of the base relations Ri (1≦i≦n) will be in a join view JV based on the base relations R1, R2, . . . , Rn. This is illustrated in
. . .
Ri
. . .
Rn) is defined conceptually (that is, JV′ is an imaginary table that does not exist in the database system). The virtual join view is a join view without the projection applied. Therefore, JV′ contains all attributes of the base relations. Each tuple in the actual join view JV comes from (and is a subset of) one tuple in JV′. Conceptually, if the tuples in JV are expanded to include all the attributes of JV′, then the tuples in JV are different from each other. After computing the change Δ=σ(Ri
. . .
ΔRi
. . .
Rn), for each tuple TA1 in Δ, there are three possible cases:
For example, suppose each tuple without projection originally has three attributes (a, b, c). After projection, the tuple only contains two attributes (a, b). Suppose there are tuples T1=(1, 2, 3), T2=(1, 2, 4), and T3=(1, 5, 6). After projection, the tuples become T1′=(1, 2), T2′=(1, 2), T3=(1, 5). Thus, the expanded tuple of T1′ is T1, the expanded tuple of T2′ is T2, and the expanded tuple of T3′ is T3. Suppose the tuple T2=(1, 2, 4) is updated to (1, 8, 5). Then, after projection, one of the tuples (1, 2) is changed to (1, 8). However, T1′ and T2′ look the same. Suppose T1′ is changed from (1, 2) to (1, 8). Note T2′ originally comes from T2, so T2′ should be changed instead of T1′. However, a distinction between T1′ and T2′ in JV cannot be made as they look the same. The tuples before the projection become T1=(1, 2, 3), T2=(1, 8, 5), and T3=(1, 5, 6). The tuples after projection become T1′=(1, 8), T2′=(1, 2), T3′=(1, 5).
However, conceptually, it can be thought that the expanded tuple of T1′ is T2, the expanded tuple of T2′ is T1, and the expanded tuple of T3′ is T3. That is, conceptually it can be thought that the expanded tuple of the identified tuple, T1′, is equal to tuple T2, even if T1′ originally comes from tuple T1. Note tuples in the relation have no order. If the order of tuples in the relation is switched, the same relation can be obtained. Thus, if the order of T1′ and T2′ is switched, the tuples after projection become T1′=(1, 2), T2′=(1, 8), T3′=(1, 5). Then it can be thought that the expanded tuple of T1′ is T1, the expanded tuple of T2′ is T2, and the expanded tuple of T3′ is T3. The key point is that tuples in the JV with the same value have no difference and thus can be treated in any way.
Thus, conceptually,
That is, the intersection of the updates to JV by transactions T and T′ is empty. Consequently, transaction T's writes to join view JV are not written by any other transaction T′ until transaction T completes (this proves part 2 of Assertion 1 for the general case).
The Y-lock locking mechanism for join views also applies to aggregate join views. An example aggregate join view AJV is created as follows:
As shown in
Consider an aggregate join view v=γ(π(σ(R1. . .
Ri
. . .
Rn))), where γ is the aggregation operator. A virtual (imaginary) join view JV=π(σ(R1
. . .
Ri
. . .
Rn)) is defined so that AJV=γ(JV). In other words, the aggregate join view AJV is equal to the virtual join view after aggregation.
An example of this is illustrated in
For purposes of proving the assertions above, the virtual or imaginary JV remains consistent with AJV. The schema of the virtual join view JV is (a, d). Aggregation is not performed on the virtual join view JV so that the join result tuples are added into new rows of the virtual join view JV. However, note that aggregation of the virtual join view JV will cause it to result in AJV.
Thus, whenever AJV is updated by Δ, conceptually, the corresponding tuples that produce Δ are updated in the virtual join view JV. By the above reasoning, if the virtual join view JV is considered instead of the aggregate join view AJV, then any parallel execution of the transactions are equivalent to some serial execution of these transactions. AJV is always set equal to γ(JV). Thus if the virtual join view JV is replaced by AJV, any parallel execution of the transactions are still equivalent to some serial execution of the transactions. This provides the four assertions above.
As noted above, a benefit of the Y-lock locking mechanism is that multiple transactions can place Y locks concurrently on the same join view JV to enhance database performance. However, in the context of projection join views and aggregate join views, a physical undo of an update of a join view due to a first transaction (T1) that aborts may not be possible. An example is illustrated in
The first transaction T1 updates the tuple (1, 1) in JV to (1, 2). The modified projection JV is shown as 152, where (1, 1) has been changed to (1, 2). In the example, another transaction T2 is also active. T2 updates tuple (1, 2) in JV to (1, 3). The modified projection JV is shown as 154.
After the T2 update, the first transaction T1 aborts. In that case, the tuple (1, 2) that was changed from (1, 1) needs to be un-updated back to (1, 1). However, that tuple has already been changed to (1, 3) by transaction T2, so a physical undo of that tuple is not feasible.
In accordance with some embodiments, a logical undo is performed. In the logical undo, the database system looks for another tuple in the join view JV that has the attribute values (1, 2). That other tuple is changed to the value (1, 1) for a logical undo of transaction T1, shown as 156 in
A logical undo is also used for aggregate join views. As shown in
If T1 aborts, a physical undo is not possible, since the entry (1, 8) no longer exists in AJV. Therefore, a logical undo is performed to change the tuple (1, 12) in AJV to (1, 9), shown as 166. This removes the contribution of the (1, 3) tuple for transaction T1 from the AJV.
By reducing the occurrences in which transactions block each other due to locks placed on a materialized view, such as a join view, database system performance is enhanced. This is especially beneficial in a parallel database system having plural data server modules, such as that shown in
IV. No-Lock Locking Mechanism
In accordance with another embodiment, instead of a Y-lock locking mechanism, a “no-lock” locking mechanism is employed for the join view JV. One purpose of placing a Y lock on a join view JV during an update of the join view in a first transaction is to prevent other transactions from reading JV, as write conflicts on the join view JV have already been handled by placing proper locks on the base relations of JV. However, in another embodiment, according to the “no-lock” locking mechanism, a Y lock (or any other type of lock) on the join view JV can be omitted while still protecting the join view JV as well as maintaining transaction serializability. A benefit of not placing any lock on the join view reduces the overhead associated with maintaining join views in response to update of base relations. Many transactions in the database system 10 are small updates that involve single-tuple or few-tuple updates to base relations of a join view. If locks (such as Y locks) are required for all such small updates, then the required locks may cause a bottleneck, since the lock manager may become tied up with many lock requests.
In some embodiments, a join view cannot be directly updated or modified. In other words, a join view JV is updated or modified by updating the base relations of the join view JV.
In the discussion below, it is assumed that a join view JV is defined on base relations R1, R2, . . . , and Rn. As shown in
If no conflicting locks on the base relations are detected, then the data server module 12 places a table-level X lock (at 408) on each base relation Rt
The table-level X lock on each base relation that is being updated can be replaced with a less restrictive SIX lock (S+IX) if the following condition is satisfied: the update of Rt
If the data server module 12 determines (at 402) that the transaction does not involve both a read and write of the join view JV, then the data server module 12 determines (at 410) if the transaction involves a read (but not a write) of JV. If so, then the data server module 12 checks (at 412) for any conflicting locks placed on the base relations R1, R2, . . . , Rn. If a conflict is present, lock conflict processing is performed (at 406). If no conflicting lock exist, then the data server module 12 places (at 414) a table-level S lock on each base relation Ri (1≦i≦n) of JV. Note that the required table-level S lock on each base relation would conflict with an X lock or IX lock placed on any of the base relations.
The requirement (by the no-lock locking mechanism) of an S lock on each base relation of JV for a transaction that reads the JV differs from that of the Y-lock locking mechanism. In the Y-lock context, to read JV, only a Y lock is needed on JV, with no S locks needed on the base relations. Therefore, for reads of JV, the Y-lock locking mechanism requires less locks than the no-lock locking mechanism.
If, however, the data server module 12 determines (at 416) that the transaction involves a write (but not a read) of the join view JV through updating one or several base relations Rt
A value-write or value-read is a write or read requested by a query with a WHERE clause condition that specifies a specific value or values for the value locking attribute of the base relation. An example query that specifies a value-write is as follows:
On the other hand, a table-write or table-read is a write or read requested by a query with a WHERE clause condition that does not specify a specific value or values for the value locking attribute of the base relation. An example query that specifies a table-read is as follows:
If h≧2, which means that more than one base relation Rt
The no-lock locking mechanism requires less locks than the Y-lock locking mechanism for transactions that cause the join view JV to be updated. In the Y-lock context, for updates of JV, a Y lock is placed on JV along with an X lock (or IX lock) on each updated base relation and an S lock on each non-updated base relation. On the other hand, in the no-lock context, for updates of JV, no lock is placed on JV, with an X lock (or IX lock or SIX lock) placed on each updated base relation and an S lock placed on each non-updated base relation.
The no-lock locking mechanism for join views can be extended to allow value locks on join views. Consider a join view JV defined on base relations Ri, R2, . . . , and Rn. For a fixed i (1≦i≦n), suppose that Ri.ai is the value locking attribute of base relation Ri that also appears in JV. Then value locking on Ri.ai for JV is allowed. However, note that no lock is placed on JV according to the no-lock locking mechanism.
If transaction T reads JV by specifying Ri.ai value(s), then transaction T places an IS lock on Ri, one or several S value locks on Ri.ai for Ri (not JV), and an S lock on each other base relation Rj (j≠i, 1≦j≦n) of JV. If transaction T both updates base relation Ri by specifying the Ri.ai value(s) and reads JV, an SIX lock is placed on Ri, one or several X value locks are placed on Ri.ai for Ri, and an S lock is placed on each other base relation Rj (j≠i, 1≦j≦n) of JV.
To show that the no-lock locking mechanism maintains the isolation properly (serializability) of transactions, the same four assertions as for the Y-lock locking mechanism are proven. It is assumed that join view JV allows duplicate tuples. If no duplicate tuples are allowed in JV, assume that each tuple in JV has a dupcnt attribute (or some other named parameter) recording the number of copies of the tuple. The following assumptions are made (the other more complex cases can be proved in a similar way): (1) any transaction T updates at most one base relation of JV; and (2) if transaction T tries to update base relation Ri (1≦i≦n), it places an S lock on each Rj (1≦j≦n, j≠i).
If transaction T writes join view JV, T places a table-level 1× or X lock on the base relation being updated until T completes. Thus transaction T's writes to join view JV are not read by any other transaction T′ until transaction T completes, since T′ requires a table-level S lock (or X lock if T′ tries to both read and update JV) on each base relation of JV. This proves the first part of Assertion 1.
In addition, if transaction T writes join view JV, there are two possible cases:
Thus transaction T's writes to join view JV are not written by any other transaction T′ until transaction T completes (this proves the second part of Assertion 1).
Based on a similar reasoning to the proof of Assertion 1, transaction T does not overwrite dirty data of other transactions in join view JV (this proves Assertion 2).
Suppose that transaction T reads join view JV by requiring a table-level S lock on each base relation of the join view JV. If some other transaction T′ writes any data in join view JV, T′ will place a table-level IX or X lock on the base relation being updated until T′ completes. Thus transaction T does not read dirty data from transaction T′ in join view JV (this proves Assertion 3). The case that transaction T tries to both read and update the join view JV is similarly proved.
If transaction T reads join view JV, T places a table-level S lock on each base relation of the join view JV until T completes. Thus no other transaction T′ can write any data in JV until transaction T completes, since T′ requires a table-level 1× or X lock on the base relation being updated (this proves Assertion 4). The case that transaction T tries to both read and update the join view JV is similarly proved.
Similar to the Y-lock locking mechanism, the no-lock locking mechanism can also be used for aggregate join views.
Since multiple transactions may also be able to update a join view JV concurrently using the no-lock locking mechanism, a logical undo mechanism similar to that discussed for the Y-lock mechanism is used in case a transaction aborts.
V. Name-Lock Locking Mechanism
According to another embodiment, instead of the Y-lock or no-lock locking mechanism discussed above, an alternative locking mechanism is a name-lock locking mechanism. As shown in
According to one embodiment, each name-lock data structure is of the following format: (NAME, COUNT). NAME refers to a base relation name, while COUNT refers to the number of transactions updating join view JV through updating base relation NAME. For join view JV defined on base relations R1, R2, . . . , and Rn, the value of NAME can be one of R1, R2, . . . , and Rn. Furthermore, NAME can have a special value (referred to as “Q”) that is not any base relation name, but indicates that a transaction is attempting to update multiple base relations of JV or that a transaction requires both a read and update of the join view JV. Note that the designation of “Q” is provided as an example only, as other designations can be used in other embodiments.
Use of the name-lock data structure is described in connection with the name-lock locking logic shown in
If the transaction does not involve a read of the join view JV, as determined at 504, the data server module determines (at 512) if the transaction involves a write to (but not a read of) the join view JV through the update of a single base relation Ri of JV. If JV is being updated through the update of a single base relation Ri, the data server module checks (at 514) if there is an S or X lock present on JV. The write of base relation Ri in the received transaction requires an X lock on Ri. Thus, if an S lock already exists, then the X lock request is blocked. Similarly, if an X lock for another transaction is already placed on Ri, then the required X lock for the received transaction will be blocked except for the case where the X lock for the other transaction is placed on the same base relation Ri. The data server module checks the name-lock data structure (NAME, COUNT) to determined if an X lock is present. COUNT equals zero means that there is no other X lock granted on JV. COUNT greater than zero indicates that there is another X lock granted on JV. However, the X lock for the received transaction can be granted if the prior X lock is on the same base relation Ri (that is, NAME=Ri). If the X lock is possible, which means that there is no S lock on Ri and that either (i) COUNT=0, or (2) COUNT>0 but NAME=Ri, then the data server module places (at 516) the X lock on JV. The name-lock data structure is then updated (at 518) by the data server module. The update is as follows: (a) if COUNT=0, then NAME is changed to Ri and COUNT is incremented by 1; or (b) if COUNT is greater than zero and NAME is equal to Ri, then COUNT is incremented by 1.
The acts performed at 512, 514, 516, and 518 are for a write to the join view JV through an update of a single base relation Ri. However, if the write to JV is a result of updating multiple base relations Ri of JV, as determined (at 520), then the data server module checks (at 522) if no S lock is present on JV and no X lock is present on JV (that is, COUNT=0). In other words, an X lock for the received transaction (which involves a write to JV through the update of multiple base relations of JV) is allowed only if there is no prior S lock or X lock on JV. If this condition is true, then the data server module places (at 524) an X lock on JV.
In addition, the name-lock data structure is updated (at 526) by changing the value of NAME to Q and incrementing COUNT by 1. As a result of this X lock (with NAME=Q), all other transactions that try to read or update JV are blocked. This is contrasted to the situation where a write to JV results from an update of a single base relation Ri of JV, in that the X lock (with NAME=Ri) does not block a subsequent transaction that tries to update JV if the subsequent transaction also involves a write to JV that results from an update of the same base relation Ri.
Another type of transaction, which is not any one of the transactions above, involves both a read and write of the join view JV. If the data server module detects (at 528) this type of transaction, then the acts at 522, 524, and 526 are performed to determine if an X lock, with NAME set to Q, can be placed on the join view JV.
When a transaction finishes execution (aborted or committed), the data server module checks (at 552) if this transaction has placed an X lock on the join view JV or not. If so, then the data server module updates (at 556) the name-lock data structure by decrementing COUNT by 1. Then the data server module releases the X lock (at 558) on the join view JV held by this transaction. Otherwise, the data server module checks (at 554) if this transaction has placed an S lock on the join view JV or not. If so, the data server module releases the S lock (at 560) on the join view held by this transaction.
With the name-lock locking mechanism, there are five possible scenarios in the database system at any time. In a first scenario, there is no transaction trying to read or update the join view JV. In a second scenario, several transaction are reading the join view JV concurrently. These transactions place table-level S locks on the join view JV, blocking other transactions from updating the base relations of JV.
In the third scenario, there are several transactions updating the same base relation Ri (1≦i≦n). These transactions all place table-level X locks on the join view JV with the parameter NAME=Ri. Provided the transactions updating R1 do not block each other without join views, they will not block each other in the presence of join views. However, such transactions will block other transactions from either reading the join view JV or updating other base relations Rj (j≠i, 1≦j≦n).
In a fourth scenario, one transaction is updating several base relations of JV simultaneously. This transaction places a table-level X lock on the join view JV, with the parameter NAME set to Q. This will block other transactions from either reading the join view JV or updating base relations R1 (1≦i≦n) of the join view. In a fifth scenario, one transaction is both reading the join view JV and updating one or several base relations of JV simultaneously. This scenario is similar to the fourth scenario.
As with the Y-lock and no-lock locking mechanisms, the name-lock locking mechanism can be extended to allow value locking on join views. Given a join view JV defined on base relations Ri, R2, . . . , Rn, and given Ri.ai is the value locking attribute of base relation R1 that also appears in JV, then X or S value locking on Ri.ai for JV is allowed. For example, consider a transaction that only updates Ri. If the update of Ri can be performed by value locking, then transaction T can also place an IX lock on JV and one or several X value locks on Ri.ai for JV (instead of a table-level X lock on JV with NAME=Ri).
To show that the name-lock locking mechanism maintains the isolation property (serializability) of transactions, for a join view JV defined on base relations R1, R2, . . . , and Rn, the same four assertions (Assertions 1–4) as for the Y-lock locking mechanism above are shown.
The four assertions are first proven for the simple case where JV=σ(R1. . .
Ri
. . .
Rn). It is assumed that join view JV allows duplicate tuples. If no duplicate tuples are allowed in JV, assume that each tuple in JV has a dupcnt attribute (or some other named parameter) recording the number of copies of that tuple.
If transaction T writes join view JV, T places a table-level X lock on JV until it completes. Thus transaction T's writes to join view JV are not read by any other transaction T′ until transaction T completes, since T′ requires a table-level S lock on JV. This proves the first part of Assertion 1.
If transaction T writes join view JV, there are three possible cases:
Suppose that transactions T and T′ update ΔRi and ΔRi′ of base relation Ri, respectively. There are three possible scenarios:
Thus transaction T's writes to join view JV are not written by any other transaction T′ until transaction T completes (proof of second part of Assertion 1).
Based on a similar reasoning to the proof of Assertion 1, transaction T does not overwrite dirty data of other transactions in join view JV (proof of Assertion 2).
Suppose that transaction T reads join view JV by requiring a table-level S lock on JV. If some other transaction T′ writes any data in join view JV, T′ places a table-level X lock on JV until T′ completes. Thus transaction T does not read dirty data from transaction T′ in join view JV, since the required S lock on JV conflicts with the X lock on JV by transaction T′ (proof of Assertion 3).
If transaction T reads join view JV, T places a table-level S lock on JV until T completes. Thus no other transaction T′ can write any data in JV until transaction T completes, since T′ requires a table-level X lock on JV. This proves Assertion 4.
In the general case where JV=π(σ(R1. . .
Ri
. . .
Rn)), the proofs of the four assertions are the same except for the case where ΔRi∩ΔRi′=≠Øin Assertion 1. For this general case, the proof is the same as the above for the Y-lock locking mechanism.
The name-lock locking mechanism for join view also applies to aggregate join view. The proof is the same as for the Y-lock locking mechanism, described above.
A first special case for which the name-lock locking mechanism can be modified to enhance parallel execution of transactions is described. Consider a join view JV AB. Suppose that the join condition of A
B is A.c=B.d, where A.c and B.d are the value locking attributes of base relations A and B, respectively. Consider the following two transactions T1 and T2. Transaction T1 inserts tuple TA1 into base relation A and tuple TB1 into base relation B. Transaction T2 inserts tuple TA2 into base relation A and tuple TB2 into base relation B. Assume that TA1.C, TB1.d, TA2.c, and TB2.d are different from each other. Then transactions T1 and T2 can be executed in parallel by changing the name-lock locking mechanism as follows:
Thus, even though each of transactions T1 and T2 is updating multiple base relations of the join view JV, the requirement of an X lock with NAME=Q can be avoided if the join condition AB is A.c=B.d and A.c and B.d are the value locking attributes of A and B, respectively. Instead of NAME=Q, the X lock placed on JV is performed with NAME=Ri (the name of the base relation being updated).
In this way, when transactions T1 and T2 do not conflict with each other without the join view JV, they also do not conflict with each other in the presence of the join view JV. The general principle for this first special case is as follows: consider a join view JV defined on base relations R1, R2, . . . , and Rn. Suppose that the join condition among base relations Rt
A second special case for which the name-lock locking mechanism can be modified to enhance performance involves the key-foreign key constraint. An example to illustrate the second special case is as follows: consider a join view JV=AB, as shown in
B is A.c=B.d, where A.c is a key of A and B.d is a foreign key of B that references A.c. Also suppose that A.c and B.e are the value locking attributes of base relations A and B, respectively. If transaction T inserts a tuple TA into base relation A, transaction T does not need to put any lock on B or JV for join view maintenance, as there can be no matching tuple in B that can be joined with TA.
A foreign key identifies table relationships. For example, table A can be a customer table, with A.c being the customer identifier and the other attributes of A describing various aspects of each customer. There is one row per customer in table A. Table B is a customer order table that contains orders of each customer, with B.d being the customer identifier. Inserting a new row into A means that a new customer (associated with a new customer identifier A.c) is being added to A. Therefore, since A.c is new, there can be no row in the customer orders table (table B) that contains order information for the new customer.
Transaction T only needs to put an IX lock on A and an X value lock on A for TA.C for the newly inserted TA. On the other hand, if transaction T inserts a tuple TB into base relation B, transaction T needs to put an IS lock on A and an S value lock on A for TB.d to do join view maintenance. Transaction T also needs to put an IX lock on B, an X value lock on B for TB.e, and a table-level X lock on JV whose NAME=B.
The general principle is as follows: consider a join view JV defined on base relations Ri, R2, . . . , and Rn. Suppose that the join condition between base relations Ri (1≦i≦n) and Rj (j≠i, 1≦j≦n) is Ri.ai=Rj.aj, where Ri.ai is a key of Ri and Rj.aj is a foreign key of Rj that references Ri.ai. Also suppose that Ri.ai is the value locking attribute of base relations Ri, and Rj.bj is the value locking attribute of base relations Rj. Then if transaction T inserts a tuple Ti into base relation Ri, transaction T requires the following locks: an IX lock on Ri and an X value lock on Ri for Ti.ai. If transaction T inserts a tuple Tj into base relation Rj, transaction T requires the following locks: an IS lock on Ri, an S value lock on Ri for Tj.aj, an IX lock on Rj, an X value lock on Rj for Tj.bj, and a table-level X lock on JV whose NAME=Rj.
A logical undo mechanism is similarly used for the name-lock locking mechanism.
VI. Auxiliary Relations
Several locking mechanisms have been described above to improve transaction concurrency in the database system by reducing the occurrences in which locks placed on a join view, and/or base relations on which the join view is based, will cause transactions to block each other from execution. According to one embodiment, auxiliary relations are also optionally used to maintain join views. The use of auxiliary relations makes join view maintenance more efficient, especially in the parallel database system shown in
Additionally, in one embodiment, a clustered index is maintained on each auxiliary relation. Index IA is maintained on attribute A.c for auxiliary relation ARA. Likewise, index IB is maintained on attribute B.x for auxiliary relation ARB. In
The auxiliary relations ARA and ARB are constructed based on a reorganization of the tuples TA and TB of base relations A and B, respectively, in which the join attributes (A.c and B.x) influence the construction. Auxiliary relations ARA and ARB thus include all the tuples of relations A and B, respectively, with the tuples simply being rearranged.
In one embodiment, the data structures of
Auxiliary relations ARA and ARB, as well as the join view JV, are also distributed across the data server modules of the parallel database system. The partitioning strategy for each of auxiliary relations ARA and ARB determines the node upon which the tuples TARA and TARB of the auxiliary relations are stored.
In an example join view definition, tuple TA of relation A is to be joined with one or more tuples TB of relation B in which attribute c of TA equals attribute x of one or more tuples TB. In one embodiment, the join operation is performed using, not relations A and B, but the auxiliary relations ARA and ARB. Then, the join result tuples are stored in the join view JV.
As noted above, a materialized join view must be maintained as base relations of the join view are modified. This means that, as each existing tuple is updated or deleted or as each new tuple is inserted into either relation A or B, the materialized view is updated so as not to become stale. The efficiency of maintenance of the materialized view depends on how the data is organized. Auxiliary relations are used to perform materialized join view maintenance, in accordance with some embodiments, to enhance efficiency.
Take, for example, a join view JV, constructed from relations A and B. If base relations A and B are partitioned on the join attributes A.c and B.x, respectively, performing the join operation is relatively efficient, since tuples TA whose “c” attributes are equal to the “x” attribute of tuples TB are stored on the same node. In the ensuing discussion, the term “node” is used broadly to refer to a data server module 12, or to a data server module plus other components. As described further below, a node can also include a parser, optimizer, scheduler, re-scheduler, and query combiner.
In a database system with multiple nodes, assume node i includes tuples TA in which attribute “c” is between 1 and 5. Also, node i includes tuples TB in which attribute “x” is between 1 and 5. Other tuples TA and TB of relations A and B are stored at nodes other than node i in the parallel DBMS. The join of (A.c=B.x) of tuples TA and TB is relatively efficient in this arrangement, since the tuples TA and TB are stored on the same node i.
Where the join view JV is also partitioned according to either attribute “c” of relation A or attribute “x” of relation B, the join result tuples also remain at node i. However, where the join view JV is not partitioned on these attributes, however, the join result tuples may need to be sent to the other node(s).
More generally, assume the join view JV is partitioned on an attribute of relation A. Incoming tuple TA is joined with the appropriate tuple or tuples of relation B at node i. The join result tuples (if any) are sent to some node k based on the attribute value of TA, as shown in
If, instead, the join view JV is not partitioned on an attribute of relation A, the distribution of join result tuples is depicted in
The situation is worse when the base relations A and B are not partitioned on the join attributes. Instead of going to a node (such as node i) in which “like” tuples will be present, multiple nodes of the parallel database system may have to be accessed to find tuples TA or TB that meet the join criteria.
In
In
Thus, as can be seen from the example of
According to one embodiment, auxiliary relations are used to overcome the shortcomings of the join view maintenance techniques described above. In general, an auxiliary relation is maintained for each relation involved in the join operation. For the general case, it is assumed that neither base relation is partitioned on the join attribute. If, however, some base relation is partitioned on the join attribute, the auxiliary relation for that base relation is unnecessary, in one embodiment.
Where a join view JV is defined on base relations A and B, two auxiliary relations, ARA for relation A, and ARB for relation B, are maintained. Relation ARA is a copy of relation A that is partitioned on the join attribute A.c. Relation ARB is a copy of relation B that is partitioned on the join attribute B.x. Additionally, as depicted in
By maintaining auxiliary relations ARA and ARB for relations A and B, respectively, assurance can be made that, at any node i of the parallel database system, tuples TA coexist with tuples TB in which the join attributes are of the same value. In other words, where the tuples TA and TB of relations A and B are not organized such that tuples meeting the condition A.c=B.x coexist at the same node, such a condition is provided using tuples TA and TB of auxiliary relations ARA and ARB.
One procedure for maintaining a join view using auxiliary relations is depicted in
In
In
Operations in which tuple TA is deleted from base relation A or updated in base relation A are similarly performed. Also, when a tuple TB is inserted into, deleted from, or updated in base relation B, similar operations are performed.
By using one or more auxiliary relations, the join view may be maintained more efficiently. In one embodiment, network communication is reduced. For each inserted (updated, deleted) tuple of base relation A, the join work to be done occurs at one node rather than at every node of the parallel database system. Further, for each inserted (updated, deleted) tuple of base relation A, the auxiliary relation ARB at one node (rather than at all nodes) is locked when performing the join operation. In one embodiment, this improves accessibility to the base relation B while join view maintenance is being performed.
In one embodiment, the storage overhead for each auxiliary relation may be kept relatively small in many cases. For example, if a join view has some selection condition on the base relation A in the where clause, such as:
As another example, a join view does not contain all attributes of the base relation A, such as:
The auxiliary relation ARA may maintain fewer than all the attributes of relation A. In the above example, the auxiliary relation ARA may maintain only the join attribute (A.c) and the attributes appearing in the select clause (A.e). Accordingly, ARA would include attributes c and e of base relation A.
Another example involves the join condition in the join view that is based on key and referential integrity restraints, such as:
Where A.c is a key of relation A and B.x is a foreign key of relation B that references to A.c. If a tuple TA is inserted into relation A, there is no matching tuple in relation B that can be joined with TA. However, if a tuple TB is inserted into relation B, there must be a tuple of relation A that can be joined with it. The case for deletion is similar. Thus, in one embodiment, if only insertion and deletion in the base relations is considered, only the auxiliary relation ARA is maintained. There is no need for auxiliary relation ARB.
A further discussion of maintaining a view with auxiliary relations is described in U.S. Ser. No. 09/900,280, filed on Jul. 6, 2001, by Gang Luo, Curt J. Ellmann, and Jeffrey F. Naughton, which is hereby incorporated by reference.
VII. Auxiliary Indices
Instead of auxiliary relations, which require relatively substantial storage overhead in some cases, auxiliary indices can be used instead. For each base relation A and B on which a join view JV is maintained, an auxiliary index AIA (for relation A) and an auxiliary index AIB (for relation B) are also maintained, as shown in
Auxiliary index AIA is an index on the join attribute A.c. AIA is partitioned on A.c. Each entry of the auxiliary index is in the form (value of A.c, list of global row identifiers), where the list of global row identifiers contains all the global row identifiers of the tuples of relation A whose attribute A.c is of that value. Each global row identifier is of the form (node identifier, local row identifier at the node). A local row identifier uniquely identifies a row of a relation at a node. A global row identifier uniquely identifies a row of a relation among all the nodes.
The auxiliary index AIA is distributed clustered if the base relation is clustered on the join attribute A.c at each node. On the other hand, the auxiliary index AIA is non-distributed clustered if the base relation is non-clustered on the join attribute at each node. An auxiliary index AIB is similarly maintained for relation B, and is of the form (value of B.x, list of global row identifiers).
As shown in
However, if JV is not partitioned on an attribute of A, then the join result tuples are distributed to plural nodes to be inserted into JV, as shown in
Generally, auxiliary indices have less storage overhead than auxiliary relations. However, join view maintenance with auxiliary indices incur more inter-node communication than for join view maintenance with auxiliary relations, but less inter-node communication than for join view maintenance without either auxiliary indices or auxiliary relations (see
The auxiliary index mechanism can also be extended to maintain a join view JV defined on more than two base relations. For optimal performance, a hybrid mechanism that includes both the auxiliary relation mechanism and the auxiliary index mechanism can be used.
VII. Locking Mechanisms for Auxiliary Relations and Auxiliary Indexes
When base relations are updated and materialized views are incrementally updated, the proper locks are placed on the base relations; there is no need to place any X or S lock on any auxiliary relation according to one embodiment. However, short-term latches on pages and locks on access paths placed by the database system are needed. In an alternative embodiment, locks can be placed on the auxiliary relation when reading or updating the auxiliary relation.
For example, in one embodiment, an X or S lock on an auxiliary relation ARR for base relation R is not needed because the auxiliary relation ARR is only updated when base relation R is updated. Therefore, the X or IX lock placed on R will prevent a subsequent conflicting update on R that causes ARR to be updated. Similarly, an auxiliary relation ARR is only read when some join view defined on base relation R is being maintained (with an S or IS lock placed on R). Thus, while a first transaction is ready for join view maintenance, the S lock or IS lock on R blocks any subsequent conflicting update of ARR through the update of R (which requires an X or IX lock on R).
Similarly, in the auxiliary index context, according to one embodiment, only proper locks are needed on base relations when base relations are updated and materialized views are being incrementally maintained. Locks on the auxiliary index are not needed. In an alternative embodiment, locks can be placed on the auxiliary index when reading or updating the auxiliary index.
Generally, for an auxiliary relation ARR of base relation R, there are three possible scenarios in the parallel database system:
Similar scenarios exist for auxiliary index AIR kept on base relation R.
A logical undo mechanism is also applied to the lock mechanism for auxiliary relations and auxiliary indices.
Viii. Example Application of No-Lock Locking Mechanism for Join View with Join View Maintenance Performed with Auxiliary Indices
Join view maintenance can be implemented using the distributed locking mechanism in a parallel database system. Consider a join view JV=AB based on the join condition A.c=B.d. Suppose neither A nor B is partitioned on the join attribute, and suppose A.e (B.f) is the value locking attribute of A (B). In this case, two auxiliary indices AIA and AIB are created, with AIB shown in
When a tuple TA is inserted into base relation A, the following operations are performed to maintain the join view JV using the no-lock locking mechanisms for JV. First, an IX lock is placed on A and an X value lock is placed on a TA.e value for A. The tuple TA is then inserted into the base relation A. Next, an IX lock is placed on AIA and an X value lock is placed on TA.C for AIA. The data structure (TA.C, (global row identifier of TA, TA.e)) is then inserted into the auxiliary index AIA.
Next, an IS lock is placed on the auxiliary index AIB and an S value lock is placed on TA.C for the auxiliary index AIB. The data server module next finds the entry in the auxiliary index AIB according to TA.c. Next, the data server module finds all the two-tuples (global row identifier of TB, value of TB.f) for those tuples TB of base relation B that match with TA. For each two-tuple (global row identifier of TB, value of TB.f), an IS lock is placed on the base relation B and an S value lock is placed on TB.f for the base relation B.
The join of tuples TA and TB is then performed, with the join result tuple inserted into the join view JV.
A benefit of the join view maintenance mechanism described above is that no table-level S or X locks are placed on A, B, or JV for join view maintenance. Only IX or IS locks and X or S value locks on A, B, AIA, and AIB are needed while they are distributed among all the nodes. Thus, no centralized node will become a bottleneck and many transactions can be executed concurrently in a parallel database system where each transaction updates one or several base relations of a join view.
IX. Transaction Rescheduling
To reduce the occurrences of deadlocks and/or lock conflicts between transactions, a transaction rescheduling mechanism is provided in the database system. The transaction rescheduling mechanism according to some embodiments is implemented in a re-scheduler (or multiple re-schedulers) 600 in the parsing engine (or multiple parsing engines) 16′, as shown in
As an example of a lock conflict, if a join view JV=AB
C, and transaction T updates base relation A, then transaction T places table-level S locks on both base relations B and C for join view maintenance. The term “update” refers to inserting data into a relation, deleting data from a relation, or changing data in the relation. As a result of the table-level S locks, no other transaction can update B or C until transaction T is committed or aborted, which creates a conflict situation until the transaction T completes.
In another example, transactions updating different base relations of a join view JV often cause deadlocks due to their lock requests on the base relations of JV. For example, consider the following two transactions T1 and T2. Transaction T1 updates base relation A and thus requires the following locks: an IX lock on A and an X value lock on some attribute value of A, a table-level S lock on B, a table-level S lock on C, and a Y lock on JV (assuming the Y-lock locking mechanism is used). Transaction T2 updates base relation B and thus requires the following locks: an IX lock on B and an X value lock on some attribute value of B, a table-level S lock on A, a table-level S lock on C, and a Y lock on JV. If transactions T1 and T2 request the locks on A and B in the following order, a deadlock will occur:
The Ix lock on A for transaction T1 blocks the S lock on A required by transaction T2, and the Ix lock on B for transaction T2 blocks the S lock on B required by transaction T1. As a result, neither transaction T1 nor transaction T2 can proceed, which is a deadlock phenomenon.
In accordance with some embodiments, the transaction rescheduling mechanism can be used to avoid some of these lock conflicts and deadlocks by reordering transactions. The transaction rescheduling mechanism is discussed first for the situation in which each transaction only reads or updates one base relation or reads one materialized view, and no materialized view is defined on other materialized view(s). The more general situation is discussed further below.
Consider a database with base relations Ri, R2, . . . , and Rn and join views JV1, JV2, . . . , and JVm. An example conflict graph of this database is defined in connection with
Effectively, the conflict graph defines a relationship among base relations that indicates conflicts among the base relations due to their association with materialized views.
For example, suppose the database contains base relations A, B, C, D, E, and F, and join views JV1=AB
C, JV2=C
D, and JV3=E
F, then an example conflict graph is as shown in
In accordance with one embodiment, the transaction rescheduling mechanism includes two queues (as shown in
After a query is submitted by a user, the parser 18 and optimizer 20 parses and optimizes the query, then generates the execution plan for the query. Thus, receiving a transaction refers to receiving the execution plan from the optimizer 20. Each execution plan contains several execution steps. A transaction is scheduled for execution once it enters the queue Q2. The scheduler 21 processes the transactions (execution plans) in Q2. If the scheduler 21 has multiple processes or threads, then multiple execution plans can be executed concurrently. The execution steps of each execution plan are sent by the scheduler to respective one or more data server modules for execution.
Thus, a transaction is “pending” while it is being processed by the re-scheduler. A pending transaction is rescheduled and submitted to the scheduler to be scheduled for execution.
The logic for rescheduling transactions, as performed by the re-scheduler 600 in the parsing engine 16′, is shown in
In a first case, a transaction T that updates base relation Ri (1≦i≦n) is “desirable” if the following is satisfied: for transactions in Q2 that update base relations RS
In a second case, a transaction T that reads base relation Ri (1≦i≦n) is “desirable” if the following is satisfied: either (a) the read operation in T can only use table-level S lock on Ri and no transaction in Q2 updates Ri, or (b) the read operation in T can use S value lock on Ri and no transaction in Q2 updates Ri by table-level X lock.
In a third case, a transaction T that reads join view JVj (1≦j≦m) is “desirable” if no transaction in Q2 updates any base relation of JVj.
If the re-scheduler 600 finds (at 606) a desirable transaction T, the transaction T is moved by the re-scheduler 600 from Q1 to Q2, where the scheduler 21 schedules the transaction for execution. Otherwise, if a desirable transaction T is not found, the re-scheduler 600 does not scan (at 610) Q1 again until either (a) some transaction leaves Q2 or (b) the number of elements in Q1 is less than N and some new transaction enters Q1. In case (b), the re-scheduler 600 only needs to check whether the new transaction is a desirable transaction or not.
Once a transaction is committed or aborted, the transaction is removed (at 612) by the scheduler 21 from Q2. In
For better performance, the re-scheduler 600 does not always start scanning Q1 from the first element of the queue when a transaction is to be rescheduled. For example, suppose that a first transaction T1 in Q1 conflicts with a second transaction T2 in Q1. After T1 is moved from Q1 to Q2 for execution, T2 becomes the first transaction in Q1 As T1 needs some time for execution, T1 will stay in Q2 for some amount of time. During this time period, T2 cannot be moved from Q1 to Q2 for execution. However, if the re-scheduler always starts scanning Q1 from the first element, T2, the rescheduling operation will be inefficient, as T2 is always not a desirable transaction during this time period (of T1 execution). To address this issue, the re-scheduler 600 maintains the first N elements of Q1 as a “round-robin” queue. Each time the re-scheduler scans Q1, the re-scheduler starts the scan with the element next to the element at which the re-scheduler previously stopped in the “round-robin” queue. If some transaction is moved from Q1 to Q2 and there are more than N elements in Q1, the (N+1)-th element of Q1 will be moved to the end of the “round-robin” queue that is N entries long.
For example, if N=4, the elements in Q1 may be rescheduled as shown in
In the example shown in
To prevent starvation of a transaction, another number M is defined. Suppose that Q1 has been scanned M times since a given transaction T became the first element in Q1 and, in each of those M scans, transaction T was not identified as a desirable transaction. If this condition is present, then each time that Q1 is subsequently scanned to find a desirable transaction T′, the desirable transaction T′ must either (i) be transaction T itself or (ii) do not conflict with transaction T. There are three possible cases:
According to one embodiment, to find the desirable transaction T quickly, the re-scheduler 600 stores an array that contains n elements (n being the number of base relations of the join view JV). The i-th (1≦i≦n) element of the array is a 4-tuple (Ei, Fi, Gi, Hi), where Ei denotes the number of transactions in Q2 that update base relation Ri by an X value lock; Fi denotes the number of transactions in Q2 that update base relation Ri by a table-level X lock; Gi denotes the number of transactions in Q2 that read base relation R1 by an S value lock; and Hi denotes the number of transactions in Q2 that read base relation R1 by a table-level S lock.
Each Ei, Fi, Gi, and Hi (1≦i≦n) is initialized to the value zero. An example is the array 620 for base relations A, B, C, D, E, and F shown in
Once a transaction T is moved from Q1 to Q2, the corresponding 4-tuple is updated as follows:
Once transaction T leaves Q2, the reverse action is performed (that is, the corresponding elements are decremented by one).
The following describes how the array of 4-tuples (Ei, Fi, Gi, Hi) is used to identify a “desirable” transaction to be moved from Q1 to Q2. Three cases are considered, with the first case involving a transaction that updates a base relation, the second case involving a transaction that reads a base relation, and the third case involving a transaction that reads a join view.
In the discussion above, it is assumed that each operation reads or updates one base relation or reads one join view. However, if some transaction T contains several operations, to move transaction T from Q1 to Q2, the re-scheduler 600 checks whether each of the operations in transaction T conflicts with any of the transactions in Q2. Once transaction T is moved from Q1 to Q2, the corresponding Action 1, 2, or 3 (discussed above) is performed for each of the operations in transaction T. There are two cases that are handled specially:
If some materialized view is defined on other materialized view(s), such a materialized view is rewritten using only base relations. For example, if join view JV1=AB and join view JV2=JV1
C, JV2 is rewritten as A
B
C. Thus there is an edge between any two of the three base relations A, B, and C in the conflict graph of the database.
As an example, consider the database shown in
As a second example, shown in B
C. Four types of transactions are used in the example: transaction updating A; transaction updating B; transaction updating C; and transaction reading JV. Assume each transaction that updates one base relation does the update by using value locking on that base relation. The different types of transactions in this example conflict with each other, so different types of transactions cannot be executed together. To enhance database system performance, the transactions are rescheduled so that transactions of the same type are executed together. Thus, as shown in
The transaction rescheduling mechanism can also be generalized to handle some special cases. The transaction rescheduling mechanism is able to handle transactions that update multiple base relations where the join condition among the base relations involves value locking attributes of the base relations. Consider a join view JVj (1≦j≦m) defined on base relations Rq
The transaction rescheduling mechanism is also able to exploit key-foreign key constraints. Consider a join view JVj (1≦j≦m) defined on base relations Rq
The transaction rescheduling mechanism is also able to handle a locking mechanism that uses value locking. Consider a join view JVj (1≦j≦m) defined on base relations Rq
Note that the transaction rescheduling mechanism uses a heuristic algorithm and is unable to completely avoid lock conflicts. For example, suppose that two transactions try to update the same tuple in the same base relation. These two transactions may coexist in Q2 after they are rescheduled to Q2 for execution. In this case, these two transactions will block each other due to lock conflicts.
To fully utilize the parallelism of a parallel database system, such as that shown in
One way to enhance parallelism of the transaction rescheduling mechanism is to partition the conflict graph of the database into groups of connected components. For each connected component group, one node is in charge of rescheduling those transactions that are related to the base relations in the connected component group. For example, the conflict graph of
If one transaction is related to several base relations that are in different connected component groups, this transaction can be rescheduled at any node that is in charge of one of these connected components. For example, if transaction T updates both base relations A and F, T can be rescheduled at either the first database system node or the second database system node. If transaction T is rescheduled at the first node, it may conflict with those transactions updating base relation F that are rescheduled at the second node. However, this will not happen frequently in an actual database. This is because in a typical database, if several base relations are frequently updated within the same transaction, there must be some relationship among these several base relations and usually a materialized view is defined on them. In this case, all these several base relations will be within the same connected component group of the conflict graph. At the database system node in charge of that connected component group, the transaction rescheduling algorithm will reschedule the transactions (including those where several base relations are updated within the same transaction) properly to maximize the throughput of transactions in the parallel database system.
X. Parallel Transaction Rescheduling
In some cases, in a parallel database system having L system nodes, one or a few of the nodes may be more heavily loaded than other system nodes in performing transaction rescheduling. As a result, this subset of the L nodes may become a bottleneck during database system execution.
Consider a join view JV=AB, where the join condition is A.c=B.d. Suppose that A.c (B.d) is not the value locking attribute of base relation A (B). Consider two kinds of transactions: transaction T1 inserts a tuple into base relation A, transaction T2 inserts a tuple into base relation B. Assume also that 40% of all the transactions in the database are T1 and another 40% of all the transactions are T2. Transaction T1 requires the following locks: an IX lock on A and an X value lock on some attribute value of A, a table-level S lock on B, and a table-level Y lock on JV. Transaction T2 requires the following locks: an IX lock on B and an X value lock on some attribute value of B, a table-level S lock on A, and a table-level Y lock on JV. The lock manager that manages the table-level locks on B resides at a single node k. As each transaction T1 requires a table-level S lock on B, and 40% of all the transactions are T1, node k will become a bottleneck to handle all the table-level S lock requests on B. Also, as each transaction T2 requires a table-level IX lock on B, and 40% of all the transactions are T2, node k will become a bottleneck to handle all the table-level IX lock requests on B. Note that the value locks for B are partitioned on all the nodes. Thus, the lock managers that manage the value locks for B will not become a bottleneck.
To address the bottleneck problem, a group query mechanism, also implemented in the parsing engine, is provided in a parallel database system, according to some embodiments. A group query is a collection of queries with redundant table-level lock requests removed. The group query mechanism is able to substantially reduce the number of table-level lock requests, which enhances database system performance.
In accordance with some embodiments, one of four choices is selected by a parsing engine based on certain criteria: (1) no transaction rescheduling is performed; (2) transaction rescheduling (as discussed in Section IX) is performed without query grouping; (3) query grouping is performed without transaction rescheduling; and (4) both transaction rescheduling and query grouping are performed. An indicator is associated with each relation upon creation of the relation, with the indicator indicating one of the four choices above.
For a given relation, no transaction rescheduling is indicated if a user or database administrator expects that transactions operating on the given relation make up only a small portion of all the transactions in the database system (e.g., less than 5%); and the transactions on the given relation are of the same type so that use of a proper locking mechanism (such as a Y-lock locking mechanism, no-lock locking mechanism, or name-lock locking mechanism for materialized views) will enable the transactions to not conflict with each other. If transactions on the given relation make up only a small portion of all transactions in the database system, then that indicates that the relation is not a “hot” relation (i.e., a relation on which a large percentage of transactions operate). Transactions are of the “same type” if the transactions perform the same operations (update or read) on the relation and the same type of lock is used by each of the relations.
For example, for a join view based on relations A and B, a transaction updating base relation A by specifying a value lock of A and another transaction updating base relation A by specifying a value lock of A can be regarded as being the same type, as they usually do not conflict with each other. On the other hand, if one transaction updates a table using a value lock, while another transaction updates the table by using a table-level lock, then such transactions may be regarded as being of different types because of conflicting locks. Similarly, if one transaction updates one base relation, while another transaction reads this base relation by using a table-level lock, then such transactions can also be considered to be of different types.
For the given base relation, transaction rescheduling (described in Section IX above) is performed (without performing query grouping) if the user or database administrator expects that most transactions of the relation will be of different types that conflict with each other and the transactions on this relation make up only a small portion of all the transactions in the database system (e.g., less than about 5%). As discussed above, parallel transaction rescheduling can be achieved by partitioning the conflict graph of the database into connected component groups to enhance rescheduling performance.
For the given relation, query grouping is performed (without performing transaction rescheduling) if the user expects that the transactions on the given relation make up a large portion (e.g., greater than 5%) of all the transactions in the database system, and most transactions on this relation are of the same type so using a proper locking mechanism (e.g., Y-lock, no-lock, or name lock) will cause them not to conflict with each other. Query grouping eliminates duplicate table-level lock requests to enhance database system performance.
As an example, consider a join view JV=AB, where the join condition is A.c=B.d and A.c (B.d) is the value locking attribute of base relation A (B). Using the Y-lock locking mechanism for join views, transactions updating base relation A by specifying the value locking attribute value of A and transactions updating base relation B by specifying the value locking attribute value of B can be regarded as of the same type. By specifying query grouping when relations A and B are created, queries updating base relation A and queries updating base relation B are combined into group queries.
For the given relation, both query grouping and transaction rescheduling are indicated if the user or database administrator expects that transactions on this relation make up a large portion of all transactions in the database system, and the transactions on this relation will be of different types that conflict with each other. In this case, query grouping is used to reduce the number of table-level lock requests so that the lock manager will not be a bottleneck. Also, the group queries are rescheduled to further decrease the likelihood of lock conflicts and/or deadlocks.
As an example, consider a join view JV that is based on relations A and B, where the join condition is A.c=B.d and A.c (B.d) is not the value locking attribute of base relation A (B). Using the Y-lock locking mechanism for join views, transactions updating base relation A by specifying the value locking attribute value of A are of the same type since they usually do not conflict with each other. Transactions updating base relation B by specifying the value locking attribute value of B are of the same type since they usually do not conflict with each other. However, transactions updating base relation A and transactions updating base relation B are of different types since they conflict with each other due to the locks they place on base relations A and B. As a result, the user may specify query grouping and transaction rescheduling when relations A and B are created. In this way, queries updating base relation A may be combined into group queries, and queries updating base relation B may be combined into group queries. These two kinds of group queries are then rescheduled using the transaction rescheduling mechanism to reduce the number of lock conflicts and/or deadlocks.
In some embodiments, transaction rescheduling does not apply to the following type of transaction: (1) this transaction contains multiple SQL statements; (2) the user first specifies begin transaction, then submits the first SQL statement; and (3) the user does not submit the second SQL statement of the transaction until the first SQL statement of the transaction finishes execution (in general, the user does not submit the next SQL statement of the transaction until the previous SQL statement finishes execution). Thus, for a given base relation on which the majority of transactions are of the type above, no transaction rescheduling is indicated.
In the ensuing discussion, the following assumptions are made. A conflict graph is defined on all the base relations in the database. The conflict graph is partitioned into connected component groups. For each connected component group, one database system node is in charge of rescheduling those transactions that are related to the base relations in the connected component group. Each database system node includes a query parser, an optimizer, a scheduler, and a re-scheduler, as shown in
Also assume that L=g×h. As shown in
As shown in
As shown in
If choice (2) is selected (at 806), then transaction rescheduling but no query grouping is selected as the course of action for the received query, and the parsing engine at node f (1≦f≦L) that the requesting client is coupled to sends (at 812) the execution plan of the query to the re-scheduler at node j (node j may be the same as or different from node f) based on which connected component group of the conflict graph the query is related to. This is illustrated in
In response to receiving the execution plan of the transaction from the parsing engine at node f, the re-scheduler at node j reschedules (at 814) the transaction and sends the transaction to the scheduler at node j (see
If choice (3) is selected (at 806), the parsing engine at node f (1≦f≦L) that the client is coupled to sends (at 820) the execution plan of the query to the query combiner (see
Also, the mapping between the SQL statement identifier of the group query GQ and the SQL statement identifiers of the individual queries p1, p2, . . . , and pt is stored (at 826) in a table WS at node kS.
The query combiner at node ks then sends (at 828) the group query to the scheduler at node ks. When the scheduler receives the execution plan of the group query, the scheduler assigns (at 830) a transaction identifier to the execution plan and executes the group query transaction (at 832). After the group query transaction finishes execution, the mapping between the SQL statement identifier of the group query GQ and the SQL statement identifiers of the individual queries p1, p2, . . . , and pt is removed (at 834) from the table Ws. Also, each client that submits the query pj (1≦j≦t) will receive a message saying that query pj is complete (the query can be either committed or aborted).
During execution of the group query GQ transaction, if the client submitting query pj (1≦j≦t) aborts query pj, then the group query GQ transaction is aborted. This means that all the queries p1, p2, . . . , and pt are aborted. In a typical parallel database system, the abort rate of individual queries is low, so aborting the group query will not happen frequently. If the group query GQ is committed, then all the individual queries p1, p2, . . . , and pt in the group query are committed.
Read queries are handled specifically when query grouping is performed. Assume that when a read query transaction is executed, the query results are first written to a temporary result space whose name is specified in the execution plan of the transaction. Then the content in the temporary result space is returned to the user session (client). At each node k5 (1≦s≦g), the query combiner keeps another table Vs with three columns (SQL statement identifier of the group query, user session, temporary result space). Note that the user session field and the temporary result space field are not needed for update transactions. Thus, if a large percentage of transactions are update transactions, using the data structure Ws (which does not contain the session and temporary result space fields) helps to conserve storage space. However, in a different implementation, the Ws and Vs structures can be combined so that the same structure is used for both update and read transactions.
An index on the table Vs is built for the SQL statement identifiers of the group queries. When several read queries are grouped into a group query at node ks, for each individual read query, the mapping between the user session and the temporary result space is stored in the table Vs. After the group query finishes execution, the mapping between the user sessions and the temporary result spaces are removed from table Vs.
The query combiner works in the following way. The query combiner keeps a query combiner queue Q, shown in
Queries in the j-th element are combined into a group query. The execution plan of the group query is sent to the scheduler for execution.
Two queries p1 and p2 are of the same type if the following conditions are satisfied:
An element of Q is of the same type as query p if each query in that element is of the same type as query p. A small query is a query that affects a relatively small number of tuples in a relation or view. A relatively small number of tuples means one or a few tuples. Large queries that affect a relatively large number of tuples are not grouped.
For example, let pR(i) denote a query that updates base relation R by specifying a value of the value locking attribute. If M2=3 (maximum size of each element of Q), the query combiner may work as follows. Originally, as shown in
Next, a query pA(2) is inserted into the first element of Q, as shown in
Next, another query PA(3) is inserted into the first element of Q, as shown in
Referring again to
As an optimization, the query combiner at node ks′ only needs to send the SQL statement identifier and the information about table-level lock requests (instead of the entire execution plan) of the group query GQ to the re-scheduler at node u. Note that the information about value X (S) locks is implied by the information about table-level IX (IS) locks. This is due to the fact that to put an X (S) value lock on an attribute value of relation R, a table-level IX (IS) lock has to be placed on R first. The execution plan of GQ is kept at node ks′. After rescheduling, the re-scheduler at node u only needs to send the SQL statement identifier (instead of the entire execution plan) of the group query GQ back to the scheduler at node ks′.
As noted above, a group query is a collection of queries where the redundant table-level lock requests are removed. Consider n queries p1, p2, . . . , and pn. Assume that the execution plan of each query pi (1≦i≦n) contains the following mi execution steps:
All the necessary table-level locks are acquired at the first execution step. In the following execution steps (2˜mi−1), proper value locks may be acquired. At the last execution step, all the previously acquired table-level locks and value locks are released.
After performing query grouping, the execution plan of the group query GQ for the n queries p1, p2, . . . , and pn contains the following m execution steps:
Execution steps 2˜m−1 of the group query GQ contain the execution steps 2˜mi−1 of each query pi (1≦i≦n) sequentially. Several execution steps of different queries can be combined into one execution step of GQ if this does not change serializability. In this way, these n queries are either all committed or all aborted. Using group query, the number of table-level lock requests can be substantially reduced.
As an example, consider a join view JV=AB. Query pi inserts a tuple TA1 into base relation A, query p2 inserts a tuple TA2 into base relation A. Suppose that tuple TA1 and tuple TA2 have different value locking attribute values. The execution plan of query pi contains the following five execution steps:
The execution plan of query p2 contains the following five execution steps:
Then the execution plan of the group query GQ for p1 and P2 may contain the following five execution steps:
Instructions of the various software routines or modules discussed herein (such as the parser, optimizer, lock manager, transaction re-scheduler, query combiner, and so forth) are stored on one or more storage devices in the corresponding systems and loaded for execution on corresponding control units or processors. The control units or processors include microprocessors, microcontrollers, processor modules or subsystems (including one or more microprocessors or microcontrollers), or other control or computing devices. As used here, a “controller” refers to hardware, software, or a combination thereof. A “controller” can refer to a single component or to plural components (whether software or hardware).
Data and instructions (of the various software routines or modules) are stored in respective storage devices, which are implemented as one or more machine-readable storage media. The storage media include different forms of memory including semiconductor memory devices such as dynamic or static random access memories (DRAMs or SRAMs), erasable and programmable read-only memories (EPROMs), electrically erasable and programmable read-only memories (EEPROMs) and flash memories; magnetic disks such as fixed, floppy and removable disks; other magnetic media including tape; and optical media such as compact disks (CDs) or digital video disks (DVDs).
The instructions of the software routines or modules are loaded or transported to each system in one of many different ways. For example, code segments including instructions stored on floppy disks, CD or DVD media, a hard disk, or transported through a network interface card, modem, or other interface device are loaded into the system and executed as corresponding software modules or layers. In the loading or transport process, data signals that are embodied in carrier waves (transmitted over telephone lines, network lines, wireless links, cables, and the like) communicate the code segments, including instructions, to the system. Such carrier waves are in the form of electrical, optical, acoustical, electromagnetic, or other types of signals.
While the invention has been disclosed with respect to a limited number of embodiments, those skilled in the art will appreciate numerous modifications and variations there from. It is intended that the appended claims cover such modifications and variations as fall within the true spirit and scope of the invention.
Number | Name | Date | Kind |
---|---|---|---|
5832508 | Sherman et al. | Nov 1998 | A |
6243778 | Fung et al. | Jun 2001 | B1 |
6339772 | Klein et al. | Jan 2002 | B1 |
6397227 | Klein et al. | May 2002 | B1 |
6453313 | Klein et al. | Sep 2002 | B1 |
6604102 | Klein et al. | Aug 2003 | B2 |
20020038313 | Klein et al. | Mar 2002 | A1 |