Method for guaranteeing processing of messages in a continuous processing system

• US 7,818,757 B1
• Filed: 07/17/2006
• Issued: 10/19/2010
• Est. Priority Date: 07/18/2005
• Status: Active Grant

• Current Assignee: Sybase Incorporated (SAP SE)
• Original Assignee: Sybase Incorporated (SAP SE)
• Inventors: Zeyliger, Alexander, Sanin, Aleksey, Tsimelzon, Mark
• Primary Examiner(s): Sough; Hyung S
• Assistant Examiner(s): DAO, TUAN C.
• Application Number:
  US11/487,796
• Time in Patent Office: 1,555 Days
• Field of Search: None
• US Class Current: 719/321
• CPC Class Codes: G06F 11/1438 Restarting or rejuvenating
• CPC Class Codes: G06F 11/1474 in transactions G06F16/20 takes precedence
• CPC Class Codes: G06F 9/546 Message passing systems or structures, e.g. queues

Abstract

A method for processing messages ensures that every message that enters a continuous processing system is fully processed in its entirety, even in the event of a failure within the system. Messages are pushed through an execution plan that includes a set of connected “primitives” that process the message, where the primitives may be distributed across a group of servers. Messages are pushed through in accordance with a protocol that does not permit unprocessed messages to be permanently lost as they travel from input sources, between primitives, and to output sources within the execution plan. The input queue, output queue, and state (if any) associated with each primitive are saved in persistent storage at select checkpoints. If a failure occurs on a server, the server, after it starts up again, restores the primitives on that server to the input queues, output queues, and states saved at the last applicable checkpoint.

Claims

1. A method for ensuring that messages in a distributed, continuous processing system are processed in their entirety, even in the event of a system failure, the method comprising:

pushing messages through a set of connected primitives that process the messages in accordance with an execution plan, where the primitives are distributed across a plurality of servers and where the messages are pushed through in accordance with protocol that does not permit unprocessed messages to be permanently lost and that enables messages to be identified uniquely over a connection,

wherein at least one primitive is associated with an input queue and an output queue, each input queue receiving an inbound message and transferring the inbound message to said at least one primitive, and each output queue;

sending an outbound message to either a different primitive or a system output, and

buffering the outbound message in the output queue associated with the at least one primitive until the sending of the outbound message is fully completed; and

for each primitive having a state, updating the state of the primitive after fully processing each received message;

for each primitive, saving in persistent storage, at select checkpoints, any input message queue, output message queue, and state associated with such primitive; and

if a failure occurs on a server in the system, restoring each primitive on such server to the input message queue, output message queue, and state, if any, saved for the primitive at the last applicable checkpoint for the primitive, where the primitive continues processing starting from such checkpoint.

2. The method of claim 1, wherein the protocol specifies that a sender of a message save a copy of the message in persistent storage until at least a recipient of the message has saved a copy of the message in persistent storage.

3. The method of claim 2, wherein the recipient of the message sends an acknowledgement to the sender after the recipient has stored the message in persistent storage.

4. The method of claim 3, wherein the recipient of the message does not send the acknowledgement until after the message has been saved at a checkpoint.

5. The method of claim 1, wherein the protocol specifies that each sender of messages assign a unique sequence number to each message sent over a connection, where messages are sent in sequence order over a connection, wherein a first unique sequence number is assigned to a given message when the given message is sent over a first connection and a second unique sequence number is assigned to the given message when the given message is sent over a second connection and wherein the first and second unique sequence numbers are different.

6. The method of claim 5, where if the sender resends a batch of messages to the recipient, the sender always resends messages within the batch in sequence order.

7. The method of claim 5, where the sequence order corresponds to the order in which the sender generates the messages it sends.

8. The method of claim 5, wherein a primitive always processes messages it receives in sequence order.

9. The method of claim 1, wherein checkpoints are not synchronized across servers.

10. The method of claim 1, where checkpoints are not synchronized across primitives on the same server.

11. The method of claim 1, where each primitive has its own checkpoint.

12. The method of claim 1, where primitives on the same server have the same checkpoints.

13. The method of claim 1, wherein, at a checkpoint for a primitive, the input message queue, output message queue, and state associated with the primitive are saved within a transaction.

14. The method of claim 1, wherein a recovered primitive eventually becomes aligned with a neighboring component in the execution plan by communicating with such neighboring component about which messages have been transmitted between the primitive and such component.

15. A system for ensuring that messages in a distributed, continuous processing system are processed in their entirety, comprising:

a continuous processing engine (CPE); and

a plurality of servers, each server comprising an input queue, a processor, an output queue, a storage and a primitive having a state, wherein each primitive is associated with the input queue, the processor, the output queue and the storage, and wherein;

the input queue is configured to, in accordance with an execution plan, receive an input message, and buffer the input message until the input message is processed,

the processor is configured to, in accordance with the execution plan, receive the input message from the input queue, process the input message, and update the state of the primitive after the input message is fully processed,

the output queue is configured to, in accordance with the execution plan, receive an output message from the processor and buffer the output message for transfer to a system output or another primitive, and

the storage is configured to store contents, at select checkpoints, of the input queue, output queue, and the state of the primitive; and

wherein the continuous processing engine (CPE) is configured to;

manage pushing through messages by each of the primitives in accordance with the execution plan, the execution plan having a protocol that does not permit unprocessed messages to be permanently lost, and

enable messages to be identified uniquely over a connection wherein if a failure occurs on a server in the plurality of servers, each primitive associated with the failed server is restored to the input message queue, output message queue, and state, if any, saved for the primitive at a last applicable checkpoint for the primitive, wherein the restored primitive is configured to continue processing starting from the last applicable checkpoint.

16. The system of claim 15, wherein the protocol specifies that a sender of a message save a copy of the message in persistent storage until at least a recipient of the message has saved a copy of the message in persistent storage.

17. The system of claim 16, wherein the recipient of the message sends an acknowledgement to the sender after the recipient has stored the message in persistent storage.

18. The system of claim 17, wherein the recipient of the message does not send the acknowledgement until after the message has been saved at a checkpoint.

19. The system of claim 15, wherein the protocol specifies that each sender of messages assign a unique sequence number to each message sent over a connection, where messages are sent in sequence order over a connection, wherein a first unique sequence number is assigned to a given message when the given message is sent over a first connection and a second unique sequence number is assigned to the given message when the given message is sent over a second connection, and wherein the first and second unique sequence numbers are different.

20. The system of claim 15, wherein checkpoints are not synchronized across servers.

21. The system of claim 15, wherein checkpoints are not synchronized across primitives on the same server.

22. The system of claim 15, wherein, at a checkpoint for a primitive, the input message queue, output message queue, and state associated with the primitive are saved within a transaction.