Distributed database management system with node failure detection

Abstract

A node failure detector for use in a distributed database that is accessed through a plurality of interconnected transactional and archival nodes. Each node is selected as an informer node that tests communications with each other node. Each informer node generates a list of suspicious nodes that is resident in one node designated as a leader node. The leader node analyzes the data from all of the informer nodes to designate each node that should be designated for removal with appropriate failover procedures.

Description

BACKGROUND

Field of the Invention

This invention generally relates to database management systems and more specifically to detecting failures during the processing of a distributed database system.

Description of Related Art

The above-identified U.S. Pat. No. 8,224,860 discloses a distributed database management system comprising a network of transactional nodes and archival nodes. Archival nodes act as storage managers for all the data in the database. Each user connects to a transactional node to perform operations on the database by generating queries for being processed at that transactional node. A given transactional node need only contain that data and metadata as required to process queries from users connected to that node. This distributed database is defmed by an array of atom classes, such as an index class, and atoms where each atom corresponds to a different instance of the class, such as index atom for a specific index. Replications or copies of a single atom may reside in multiple nodes wherein the atom copy in a given node is processed in that node.

In an implementation of such a distributed database asynchronous messages transfer among the different nodes to maintain the integrity of the database in a consistent and a concurrent state. Specifically each node in the database network has a unique communication path to every other node. When one node generates a message involving a specific atom, that message may be sent to every node that contains a replication of that specific atom. Each node generates these messages independently of other nodes. So at any given instant multiple nodes will contain copies of a given atom and different nodes may be at various stages of processing that atom. As the operations in different nodes are not synchronized it is important that the database be in a consistent and concurrent state at all times.

A major characteristic of such distributed databases is that all nodes be in communication with each other at all times so the database is completely connected. If a communications break occurs, the database is not considered to be connected. One or more nodes must be identified and may be removed from the network in an orderly manner. Such identification and removal must consider that any node can fail at any given time and that a communications break can occur only between two nodes or that multiple breaks can occur among several nodes. The identification of a node or nodes for removal must be accomplished in a reliable manner. Moreover such an identification should enable failure processes to resolve a failure with minimal interruption to users.

SUMMARY

Therefore it is an object of this invention to provide a method for detecting a node failure in a distributed database management system.

Another object of this invention is to provide a method for detecting a node failure and for designating a node for failure.

Still another object of this invention is to provide a method for detecting a node failure and for designating a node for failure on a reliable basis.

Yet another object of this invention is to provide a method for detecting a node failure and for designating a node for failure with minimal interruption to users.

BRIEF DESCRIPTIONS OF THE DRAWINGS

The appended claims particularly point out and distinctly claim the subject matter of this invention. The various objects, advantages and novel features of this invention will be more fully apparent from a reading of the following detailed description in conjunction with the accompanying drawings in which like reference numerals refer to like parts, and in which:

FIG. 1 is a diagram in schematic form of one embodiment of an elastic, scalable, on-demand, distributed database management system that incorporates this invention and that includes interconnected transactional and archival nodes;

FIG. 2 depicts the organization of a transactional node;

FIG. 3 depicts the organization of an archival node;

FIG. 4 depicts the syntax of an exemplary asynchronous message that is transferred among the transaction and archival nodes of the system of FIG. 1.

FIG. 5 depicts certain asynchronous messages that are specifically involved with the implementation of this invention;

FIG. 6 is a flow diagram of a node failure system that incorporates this invention;

FIG. 7 is a flow diagram of a node failure detection system that identifies possible failures and that is useful in implementing the node failure system of FIG. 6;

FIG. 8 is a flow diagram of an operation that processes the information obtained in the node failure detection system of FIG. 7; and

FIG. 9 is a flow diagram of a process for choosing a node to fail based upon information obtained from the operation disclosed in FIG. 8.

DETAILED DESCRIPTION

FIG. 1 depicts one embodiment of an elastic, scalable, on-demand, distributed database management system 30 with a plurality of data processing nodes that incorporates this invention. Nodes N1 through N6 are “transactional nodes” that provide user applications access to the database; nodes A1 and A2, “archival nodes” that function to maintain a disk archive of the entire database at each archival node. While an archival node normally stores the entire database, a single transactional node contains only that portion of the database it determines to be necessary to support transactions being performed at that node at that time.

Each node in FIG. 1 can communicate directly with every other node in the system 30 through a database system network 31. For example, node N1 can establish a communications path with each of nodes N2 through N6, A1 and A2. Communications between any two nodes is by way of serialized messages. In a preferred embodiment, the messaging is performed in an asynchronous manner to maximize the bandwidth used by the system thereby to perform various operations in a timely and prompt manner. Typically the database system network 31 operates with a combination of high-bandwidth, low-latency paths (e.g., an Ethernet network) and high-bandwidth, high-latency paths (e.g., a WAN network). Each node has the capability to restrict use of a low-latency path to time-critical communications (e.g., fetching an atom). The high-latency path can be used for non-critical communications (e.g. a request to update information for a table). Also and preferably, the data processing network of this invention incorporates a messaging protocol, such as the Transmission Control Protocol (TCP) that assures that each node processes messages in the same sequence in which they were sent to it by other nodes.

FIG. 2 depicts a representative transactional node 32 that links to the database system network 31 and various end users 33. The transactional node 32 includes a central processing system (CP) 34 that communicates with the database system network 31 through a network interface 35 and with the various users through and user network interface 37. The central processing system 34 also interacts with RAM memory 38 that contains a copy of the database management program that implements a preferred embodiment of this invention. This program functions to provide a remote interface 40, a database request engine 41 and a set 42 of classes or objects.

The database request engine 41 only exists on transactional nodes and is the interface between the high-level input and output commands at the user level and system level input and output commands at the system level. In general terms, its database request engine parses, compiles and optimizes user queries such as SQL queries into commands that are interpreted by the various classes or objects in the set 42. The classes/objects set 42 is divided into a subset 43 of “atom classes,” a subset 44 of “message classes” and a subset 45 of “helper classes.”

Referring to FIG. 3, each archival node 50, such as archival node A1 or A2 in FIG. 1, also connects to the database system network 31. However, in place of end users 33 associated with a transactional node 32 in FIG. 2, an archival node connects only to persistent storage 51, typically a disk-based storage system or a key value store. The archival node 50 includes a central processing system 54 that communicates with the persistent storage 51 through an I/O channel 52 and with the database system network 31 through a network interface 55. The central processing system 54 also interacts with RAM memory 57 that contains a set 62 of classes or objects. Similarly to the transactional node 32 in FIG. 2, the classes/objects set 62 in FIG. 3 includes a set 63 of “atom classes,” a set 64 of “message classes” and a set 65 of “helper classes.”

The atom classes collectively defme atoms that contain fragments of the database including metadata and data. Atoms are replicated from one node to another so that users at different nodes are dealing with consistent and concurrent versions of the database, even during actions taken to modify the information in any atom. At any given time there is no guarantee that all replicated copies of a single atom will be in the same state because there is no guarantee that messages from one node to other nodes will be processed concurrently.

As previously indicated, communications between any two nodes is by way of serialized messages which are transmitted asynchronously using the TCP or other protocol with controls to maintain message sequences. FIG. 4 depicts the basic syntax of a typical message 90 that includes a variable length header 91 and a variable length body 92. The header 91 includes a message identifier code 93 that specifies the message and its function. The header 91 also includes a software version identification 94, a local identification 95 of the sender and information 96 for the destination of the message as an added identification. From this information the recipient node can de-serialize, decode and process the message.

FIG. 5 depicts a subset of messages having the syntax of FIG. 4 for implementing this invention. As discussed previously, when a message is to be sent, there are different communications paths to different nodes. For example, if one node requests an existing atom, replications of that atom may be obtained from a number of other nodes. In this embodiment, a known “pinging” operation one node sends a “Ping” message to a receiving node. The receiving node returns a “Ping Acknowledge” message to the sending node with time information. In one embodiment each node periodically uses a helper's class to send a Ping message to each of the other nodes to which it connects. Each receiving node uses a helper class to return a Ping Acknowledge message 111 that contains the ping time.

The failure to receive a ping acknowledgement within a predetermined time indicates a break in communications with respect to messages being sent between the requesting or sending node and the recipient or receiving node. In the context of this invention a first node transmits the Ping message 110 to another node and functions as an “informer node” or I-Node if the Ping Acknowledge signal is not received. If there is a failure, the I-Node identifies the receiving node as being suspicious (e.g. an “S-Node”) by means of a Suspicious Node message 159. A Leader Acknowledge message 160 triggers a request for each I-Node to respond with information about any suspicious nodes that connect to that specific I-Node.

A purpose of this invention is to detect a node failure and enable corrective action. FIG. 6 depicts an overall failure system 200 that includes a failure monitor 201 for detecting various types of failure. A node failure detection process 202 incorporates this invention to provide such an indication. If a failure is detected, step 203 diverts the failure system to a detected failure process 204.

Referring to FIG. 7, during an iterative node failure detection process 202, each node, generally in sequence, acts as a sending node of I-node that uses step 211 to select a receiving node and step 212 to ping the selected receiving node. For example, when N1 acts as an I-Node, steps 211 and 212 select an initial receiving node to receive a Ping message at step 212. During successive iterations the I-node will ping nodes N2 through N6, A1 and A2 using step 211 in successive iterations to select each selected receiving node in sequence. If a Ping Acknowledge message is received in a timely fashion at step 213, step 214 selects a next node, such as node N3, for receiving a Ping message.

If no Ping Acknowledge message is received within a defmed time interval, it is assumed that a communications break exists. Step 215 marks that receiving node as being “suspicious” with respect to the sending I-Node. Step 216 sends a Suspicious Node message 159 that includes the I-Node identification and Suspicious Node identification in 216 to a Leader Node.

A Leader Node is responsible for analyzing the information received from all the I-Nodes. Only one L-Node can exist at a time and it must be a non-suspicious node. A node can only act as an L-Node if it has received a response for a given set of S-Nodes from a majority of the database as represented by other I-Nodes (i.e., majority of active, non-suspicious nodes). If the active I-Node receives a Leader Acknowledgement message in a timely manner, step 217 uses step 214 to select a next node to be tested by the I-node and control returns to step 212. Otherwise, there is no certainty as to whether the I-node or the node being tested is causing the communications break. Step 220 selects the next non-suspicious node as the new Leader Node. If it is available, step 221 returns control to step 216 and the message is resent to the new L-Node. If no node qualifies as an L-Node, an error state is generated in step 222.

With respect to the process shown in FIG. 8, whenever a Leader Node receives a suspicious node message that identifies both an I-Node and a suspicious node at step 251, step 252 establishes a first time interval. Step 253 identifies the number of I-Nodes that identify the reported Suspicious Node. If a majority of the active nodes identify the Suspicious Node, step 254 immediately transfers control to a Choose Node to Fail process 255 that chooses the node that is to fail. Step 256 sends a Leader Node Acknowledge message to all I-Nodes. Control then returns to step 251 to await receipt of a the next Suspicious Node message.

If a majority does not exist at the instant of step 253, step 254 transfers control to step 257 that times out the first time interval. If the majority is reached prior to the expiration of that time interval, step 257 diverts control to steps 255 and 256. Otherwise step 257 transfers control to step 260 that sends a Leader Acknowledge message to all I-Nodes and then waits in step 261 for a second time interval to determine whether a majority of I-Nodes respond. At the end of that interval control passes through step 262 to step 255 if a majority of I-Nodes has been identified. If the second time interval expires without obtaining a majority, step 262 diverts to establish an error state at step 263.

FIG. 9 depicts the process 204 in FIG. 6 for designating nodes that should fail. The process selects one I-Node at step 271 that is paired with a Suspicious Node in step 272. Then the system processes step 273 to determine whether the selected Suspicious Node is only suspicious to the selected I-Node. If only the selected I-Node identifies a Suspicious Node, a standoff exists and step 274 diverts control to step 276 that designates the node with the highest node ID as being disabled. If the standoff does not exist, step 274 transfers control to step 275 that designates all the suspicious nodes associated with the selected I-Node to be disabled.

After either step 275, step 276 completes its process and control returns to step 271 to execute the process for another I-Node. When completed all the I-Nodes have been processed the node designations are made available to the failure system 200 in FIG. 6 for being processed in step 204.

Therefore there has been disclosed an embodiment of this invention wherein each node operating with a distributed database monitors communication paths with all other nodes in the network. Any communications break is noted and the network is analyzed to determine nodes that need to be failed. This information is reported for processing whereby failures are handled in orderly fashion with minimal interruption to user activities and in a manner in which data remains consistent and concurrent. More specifically, this invention enhances user access because it detects node failure in an orderly and efficient manner to enable appropriate failure system to maintain the database in a concurrent and consistent manner.

It will be apparent that many modifications can be made to the disclosed apparatus without departing from the invention. For example, this invention has been described with a “majority” defmed as a majority of a subset of all the active nodes. In other applications, the “majority” might be defmed as a majority of the archival nodes. Still other subsets of nodes could be defmed as the basis for determining a majority. Specific messages and processes have been given different titles for purposes of explanation. It will be apparent that equivalent messages and processes could be designed by different names while still performing the same or equivalent functions. Therefore, it is the intent of the appended claims to cover all such variations and modifications as come within the true spirit and scope of this invention.

Claims

1. A method for processing information obtained by a node failure detection system, the node failure detection system included in a distributed database, the distributed database comprising a plurality of nodes, the plurality of nodes comprising a leader node and a plurality of informer nodes, the method comprising: at each informer node in the plurality of informer nodes: transmitting a ping message to each other node in the plurality of nodes;monitoring responses to the ping message from each other node in the plurality of nodes; andresponding to an invalid response from a responding node in the plurality of nodes by designating the responding node as a suspicious node; andtransmitting a message to the leader node, the message comprising anidentification of the informer node and the suspicious node; andat the leader node: receiving the message comprising the identification of the informer node and the suspicious node;determining a number of the plurality of informer nodes that received invalid responses from the suspicious node;sending an acknowledgement message to the plurality of informer nodes if the number is fewer than a majority of the plurality of informer nodes; anddesignating the suspicious node as failed if the majority of the plurality of informer nodes identify the suspicious node in a message or the majority of the plurality of informer nodes identify the suspicious node in response to the acknowledgment message.

2. The method of claim 1, further comprising, if the leader node becomes a suspicious node, designating another node in the plurality of nodes as a new leader node.

3. The method of claim 1, further comprising: if the number of the plurality of informer nodes that received invalid responses from the suspicious node is less than the majority, waiting for a first predefined time interval to send the acknowledgement message.

4. The method of claim 1, wherein designating the suspicious node as failed if the majority of the plurality of informer nodes identifies the suspicious node in response to the acknowledgement message includes waiting for a second predefined time interval for the majority of the plurality of informer nodes to respond to the acknowledgment message.

5. The method of claim 4, further comprising: generating an error state message at the leader node if the majority of the plurality of informer nodes does not respond to the acknowledgment message within the second predefined time interval.

6. The method of claim 1, wherein each informer node in the plurality of informer nodes transmits the ping message to each other node in the plurality of nodes.

7. The method of claim 1, wherein the leader node analyzes information from each informer node in the plurality of informer nodes.

8. A method for choosing a node in a distributed database to fail, the distributed database comprising a plurality of nodes, the plurality of nodes comprising a leader node and a plurality of informer nodes, the method comprising: selecting a first informer node from the plurality of informer nodes;designating, by the first informer node, a first node in the plurality of nodes as a first suspicious node in response to an invalid response from the first node;determining if the first suspicious node is suspicious to only the first informer node;if the first suspicious node is suspicious to only the first informer node, designating the first suspicious node or the first informer node as disabled based on a higher node identification; andif the first suspicious node is suspicious to at least one other informer node in the plurality of informer nodes, designating all suspicious nodes identified by the first informer node as failed.

9. The method of claim 8, further comprising: selecting a second informer node from the plurality of informer nodes;designating, by the second informer node, a second node in the plurality of nodes as a second suspicious node in response to an invalid response from the second node;determining if the second suspicious node is suspicious to only the second informer node;if the second suspicious node is suspicious to only the second informer node, designating the second suspicious node or the second informer node as disabled based on another higher node identification; andif the second suspicious node is suspicious to at least one other informer node in the plurality of informer nodes, designating all suspicious nodes identified by the second informer node as failed.

10. The method of claim 8, wherein the leader node is a non-suspicious node.

11. The method of claim 8, wherein the distributed database includes only one leader node at any given time.