SYSTEMS AND METHODS FOR FAULT TOLERANT, ADAPTIVE EXECUTION OF ARBITRARY QUERIES AT LOW LATENCY

• US 20140108861A1
• Filed: 10/15/2013
• Published: 04/17/2014
• Est. Priority Date: 10/15/2012
• Status: Active Grant

• Current Assignee: Teradata US, Inc. (Teradata Corporation)
• Original Assignee: Hadapt Inc. (Teradata Corporation)
• Inventors: Abadi, Daniel, Bajda-Pawlikowski, Kamil, Schlussel, Rebecca, Wickline, Philip
• Granted Patent: US 9,934,276 B2
• Time in Patent Office: Days
• US Class Current: 714/11
• CPC Class Codes: G06F 11/2028 eliminating a faulty processor or activating a spare
• CPC Class Codes: G06F 16/24532 of parallel queries
• CPC Class Codes: G06F 16/24542 Plan optimisation
• CPC Class Codes: G06F 9/5066 Algorithms for mapping a plurality of inter-dependent sub-tasks onto a plurality of physical CPUs mappping at compile time, see G06F8/451

Abstract

A system and method for performing distributed execution of database queries includes a query server that receives a query to be executed on a database, forms a query plan based on the query, assigns tasks to task slots on a plurality of worker nodes in a cluster, and, upon receipt of a notification that a task has completed on a worker node, immediately assigns an unassigned task to a free task slot on that worker node, such that the task may begin executing on that worker node substantially immediately thereafter. The task slots on worker nodes include pools of resources that run tasks without start-up overhead.

Claims

1. A system for performing distributed execution of database queries, the system comprising:

a query server comprising at least one memory storing computer executable instructions and at least one processing unit for executing the instructions, wherein execution of the instructions causes the at least one processing unit to;

receive a query to be executed on a database;

form a query plan based on the query, the query plan comprising a plurality of operators, each operator divided into one or more tasks;

assign tasks to task slots on a plurality of worker nodes in a cluster, wherein the task slots comprise pools of resources that run tasks without start-up overhead; and

upon receipt of a notification that a task has completed on a worker node, immediately assign an unassigned task to a free task slot on that worker node, such that the task may begin executing on that worker node substantially immediately thereafter.

2. The system of claim 1, wherein the database comprises at least one table partitioned over the plurality of worker nodes.

3. The system of claim 1, wherein execution of the instructions further causes the at least one processing unit to:

receive a task complete status update from a worker node upon completion of an assigned task on that worker node; and

in response to the task complete status update, notify a worker node having an assigned task dependent on output of the completed task that the output may be retrieved from the worker node that completed the task.

4. The system of claim 1, wherein execution of the instructions further causes the at least one processing unit to maintain a list of task slots existing across worker nodes in the cluster, the list of task slots comprising the status of each task slot.

5. The system of claim 4, wherein the status of each task slot is selected from the group consisting of busy, free, and blacklisted.

6. The system of claim 1, wherein execution of the instructions further causes the at least one processing unit to maintain a list of tasks for the query plan, the list of tasks comprising the status of each task.

7. The system of claim 6, wherein the status of each task is selected from the group consisting of completed, scheduled, and not yet scheduled.

8. The system of claim 1, wherein execution of the instructions further causes the at least one processing unit to assign a same task to a free task slot on two or more different worker nodes, thereby causing the same task to be redundantly executed.

9. The system of claim 1, wherein execution of the instructions further causes the at least one processing unit to, upon a determination that a worker node has failed, blacklist the failed node such that no further tasks are assigned to task slots on the failed node.

10. The system of claim 9, wherein execution of the instructions further causes the at least one processing unit to reassign tasks previously assigned to task slots on the failed node to task slots on one or more other worker nodes.

11. The system of claim 1, wherein the query comprises SQL.

12. The system of claim 1, wherein the query plan comprises an operator directed acyclic graph, wherein each operator in the graph comprises a database operation to be performed on data input to that operator, and wherein each parent operator in the graph is dependent upon data produced by one or more child operators.

13. The system of claim 12, wherein execution of the instructions further causes the at least one processing unit to divide an operator in the query plan into a plurality of tasks, each task for applying the operation associated with the operator to a disjoint subset of data from the database.

14. The system of claim 13, wherein at least one of the tasks is a child task having at least one parent task dependent on subsets of output data produced by the child task.

15. The system of claim 14, wherein execution of the instructions further causes the at least one processing unit to associate annotations with the child task, the annotations comprising a mapping function for allocating the subsets of output data to the at least one parent task.

16. The system of claim 12, wherein the operators comprise one or more of scan, filter, sort, join, comparison, logical, and user-defined operators.

17. A system for performing distributed execution of database queries, the system comprising:

a plurality of worker nodes, each worker node comprising;

a first memory storing at least a portion of a database;

a second memory storing computer executable instructions; and

at least one processing unit for executing the instructions, wherein execution of the instructions causes the worker node to;

receive, from a query server, an assignment of a task to be executed on a free task slot of the worker node, wherein the task is assigned to the task slot by the query server and received by the worker node immediately after the task slot becomes available;

receive a notification, from the query server, that input required by a task assigned to a task slot of the worker node is available; and

in response to the notification, retrieve the required input from a worker node that produced the required input.

18. The system of claim 17, wherein each worker node comprises a plurality of task slots, each task slot comprising a pool of resources that runs tasks without start-up overhead.

19. The system of claim 17, wherein an assigned task on a worker node retrieves required input from one of local storage on the worker node and storage on a different worker node.

20. The system of claim 17, wherein a worker node completing a task retains output created by the completed task until execution of a database query is completed.

31. The method of claim 20, wherein the query comprises SQL.

32. The method of claim 20, wherein the query plan comprises an operator directed acyclic graph, wherein each operator in the graph comprises a database operation to be performed on data input to that operator, and wherein each parent operator in the graph is dependent upon data produced by one or more child operators.

33. The method of claim 32, wherein the execution of the instructions further results in the query server performing the step of dividing an operator in the query plan into a plurality of tasks, each task for applying the operation associated with the operator to a disjoint subset of data from the database.

34. The method of claim 33, wherein at least one of the tasks is a child task having at least one parent task dependent on subsets of output data produced by the child task.

35. The method of claim 34, wherein the execution of the instructions further results in the query server performing the step of associating annotations with the child task, the annotations comprising a mapping function for allocating the subsets of output data to the at least one parent task.

36. The method of claim 32, wherein the operators comprise one or more of scan, filter, sort, join, comparison, logical, and user-defined operators.

21. In a system comprising a query server having at least one memory storing computer executable instructions, a method for performing distributed execution of database queries, the method comprising:

executing the instructions by at least one processing unit, the execution of the instructions resulting in the query server performing the steps of;

receiving a query to be executed on a database;

forming a query plan based on the query, the query plan comprising a plurality of operators, each operator divided into one or more tasks;

assigning tasks to task slots on a plurality of worker nodes in a cluster, wherein the task slots comprise pools of resources that run tasks without start-up overhead; and

upon receiving a notification that a task has completed on a worker node, immediately assigning an unassigned task to a free task slot on that worker node, such that the task may begin executing on that worker node substantially immediately thereafter.

22. The method of claim 21, wherein the database comprises at least one table partitioned over the plurality of worker nodes.

23. The method of claim 21, wherein the execution of the instructions further results in the query server performing the steps of:

receiving a task complete status update from a worker node upon completion of an assigned task on that worker node; and

in response to the task complete status update, notifying a worker node having an assigned task dependent on output of the completed task that the output may be retrieved from the worker node that completed the task.

24. The method of claim 21, wherein the execution of the instructions further results in the query server performing the step of maintaining a list of task slots existing across worker nodes in the cluster, the list of task slots comprising the status of each task slot.

25. The method of claim 24, wherein the status of each task slot is selected from the group consisting of busy, free, and blacklisted.

26. The method of claim 21, wherein the execution of the instructions further results in the query server performing the step of maintaining a list of tasks for the query plan, the list of tasks comprising the status of each task.

27. The method of claim 26, wherein the status of each task is selected from the group consisting of completed, scheduled, and not yet scheduled.

28. The method of claim 21, wherein the execution of the instructions further results in the query server performing the step of assigning a same task to a free task slot on two or more different worker nodes, thereby causing the same task to be redundantly executed.

29. The method of claim 21, wherein the execution of the instructions further results in the query server performing the step of, upon determining that a worker node has failed, blacklisting the failed node such that no further tasks are assigned to task slots on the failed node.

30. The method of claim 29, wherein the execution of the instructions further results in the query server performing the step of reassigning tasks previously assigned to task slots on the failed node to task slots on one or more other worker nodes.

37. In a system comprising a plurality of worker nodes, each worker node having a first memory storing at least a portion of a database and a second memory storing computer executable instructions, a method for performing distributed execution of database queries, the method comprising:

executing the instructions by at least one processing unit, the execution of the instructions resulting in a worker node performing the steps of;

receiving, from a query server, an assignment of a task to be executed on a free task slot of the worker node, wherein the task is assigned to the task slot by the query server and received by the worker node immediately after the task slot becomes available;

receiving a notification, from the query server, that input required by a task assigned to a task slot of the worker node is available; and

in response to the notification, retrieving the required input from a worker node that produced the required input.

38. The method of claim 37, wherein each worker node comprises a plurality of task slots, each task slot comprising a pool of resources that runs tasks without start-up overhead.

39. The method of claim 37, wherein the execution of the instructions further results in the worker node performing the step of retrieving required input for an assigned task from one of local storage on the worker node and storage on a different worker node.

40. The method of claim 37, wherein the execution of the instructions further results in the worker node performing the step of retaining output created by the completed task until execution of a database query is completed.