Set optimizer.force-single-node-output to false by default by arhimondr · Pull Request #13217 · trinodb/trino

arhimondr · 2022-07-18T22:49:07Z

Description

With this setting enabled the optimizer will insert an additional exchange
to make sure the coordinator always consumes query results from a single
task. The exchange client used by the coordinator can consume results from
as many tasks and stages as needed. The setting was introduced in 2017
by 6143485
as a transitional setting to migrate from introducing an additional
exchange but has never been set to false by default.

Is this change a fix, improvement, new feature, refactoring, or other?

Improvement

Is this a change to the core query engine, a connector, client library, or the SPI interfaces? (be specific)

Core engine

How would you describe this change to a non-technical end user or system administrator?

Non user visiblle

Related issues, pull requests, and links

6143485

Documentation

(x) No documentation is needed.
( ) Sufficient documentation is included in this PR.
( ) Documentation PR is available with #prnumber.
( ) Documentation issue #issuenumber is filed, and can be handled later.

Release notes

( ) No release notes entries required.
(x) Release notes entries required with the following suggested text:

# Core
* Set optimizer.force-single-node-output to false by default

Stage execution could remain in the same SCHEDULING state while already running some tasks and not transition the state further until more tasks are scheduled.

With this setting enabled the optimizer will insert an additional exchange to make sure the coordinator always consumes query results from a single task. The exchange client used by the coordinator can consume results from as many tasks and stages as needed. The setting was introduced in 2017 by trinodb@6143485 as a transitional setting to migrate from introducing an additional exchange but has never been set to false by default.

sopel39 · 2022-07-19T09:36:27Z

Will it cause coordinator to perform actual computations (e.g. final aggregations?). If so, then IMO we should keep it as is.

findepi · 2022-07-19T11:03:58Z

Will it cause coordinator to perform actual computations (e.g. final aggregations?). If so, them IMO we should keep it as is.

I agree.

@arhimondr can you verify that after the change, no computations (aggregation, non-identity projection, filter, etc.) get scheduled on the coordinator?

arhimondr · 2022-07-19T16:12:32Z

@sopel39 @findepi When the optimizer.force-single-node-output is set to true the optimizer ensures that query results are consumed always from a single node. For example:

trino:tiny> EXPLAIN (TYPE DISTRIBUTED) SELECT * FROM nation;
                                           Query Plan                                            
-------------------------------------------------------------------------------------------------
 Fragment 0 [SINGLE]                                                                             
     Output layout: [nationkey, name, regionkey, comment]                                        
     Output partitioning: SINGLE []                                                              
     Output[columnNames = [nationkey, name, regionkey, comment]]                                 
     │   Layout: [nationkey:bigint, name:varchar(25), regionkey:bigint, comment:varchar(152)]    
     │   Estimates: {rows: 25 (2.67kB), cpu: 2.67k, memory: 0B, network: 2.67kB}                 
     └─ RemoteSource[sourceFragmentIds = [1]]                                                    
            Layout: [nationkey:bigint, name:varchar(25), regionkey:bigint, comment:varchar(152)] 
            Estimates:                                                                           
                                                                                                 
 Fragment 1 [SOURCE]                                                                             
     Output layout: [nationkey, name, regionkey, comment]                                        
     Output partitioning: SINGLE []                                                              
     TableScan[table = tpch:tiny:nation]                                                         
         Layout: [nationkey:bigint, name:varchar(25), regionkey:bigint, comment:varchar(152)]    
         Estimates: {rows: 25 (2.67kB), cpu: 2.67k, memory: 0B, network: 0B}                     
         nationkey := tpch:nationkey                                                             
         regionkey := tpch:regionkey                                                             
         name := tpch:name                                                                       
         comment := tpch:comment                                                                 
                                                                                                 
                                                                                                 
(1 row)

However coordinator can consume query results from any number of nodes with no issues. Here's how the plan looks like for an identical query when optimizer.force-single-node-output is set to false:

trino:tiny> EXPLAIN (TYPE DISTRIBUTED) SELECT * FROM nation;
                                           Query Plan                                            
-------------------------------------------------------------------------------------------------
 Fragment 0 [SOURCE]                                                                             
     Output layout: [nationkey, name, regionkey, comment]                                        
     Output partitioning: SINGLE []                                                              
     Output[columnNames = [nationkey, name, regionkey, comment]]                                 
     │   Layout: [nationkey:bigint, name:varchar(25), regionkey:bigint, comment:varchar(152)]    
     │   Estimates: {rows: 25 (2.67kB), cpu: 2.67k, memory: 0B, network: 0B}                     
     └─ TableScan[table = tpch:tiny:nation]                                                      
            Layout: [nationkey:bigint, name:varchar(25), regionkey:bigint, comment:varchar(152)] 
            Estimates: {rows: 25 (2.67kB), cpu: 2.67k, memory: 0B, network: 0B}                  
            nationkey := tpch:nationkey                                                          
            regionkey := tpch:regionkey                                                          
            name := tpch:name                                                                    
            comment := tpch:comment                                                              
                                                                                                 
                                                                                                 
(1 row)

For queries with the top most stage of SINGLE distribution (such as an aggregation) the plan remains unchanged. Here is the plan when the optimizer.force-single-node-output is set to true:

trino:tiny> EXPLAIN (TYPE DISTRIBUTED) SELECT count(*) FROM nation;
                               Query Plan                               
------------------------------------------------------------------------
 Fragment 0 [SINGLE]                                                    
     Output layout: [count]                                             
     Output partitioning: SINGLE []                                     
     Output[columnNames = [_col0]]                                      
     │   Layout: [count:bigint]                                         
     │   Estimates: {rows: ? (?), cpu: ?, memory: ?, network: ?}        
     │   _col0 := count                                                 
     └─ Aggregate[type = FINAL]                                         
        │   Layout: [count:bigint]                                      
        │   Estimates: {rows: ? (?), cpu: ?, memory: ?, network: ?}     
        │   count := count("count_0")                                   
        └─ LocalExchange[partitioning = SINGLE]                         
           │   Layout: [count_0:bigint]                                 
           │   Estimates: {rows: ? (?), cpu: ?, memory: ?, network: ?}  
           └─ RemoteSource[sourceFragmentIds = [1]]                     
                  Layout: [count_0:bigint]                              
                  Estimates:                                            
                                                                        
 Fragment 1 [SOURCE]                                                    
     Output layout: [count_0]                                           
     Output partitioning: SINGLE []                                     
     Aggregate[type = PARTIAL]                                          
     │   Layout: [count_0:bigint]                                       
     │   Estimates:                                                     
     │   count_0 := count(*)                                            
     └─ TableScan[table = tpch:tiny:nation]                             
            Layout: []                                                  
            Estimates: {rows: 25 (0B), cpu: 0, memory: 0B, network: 0B} 
                                                                        
                                                                        
(1 row)

Here is the plan when the optimizer.force-single-node-output is set to false:

trino:tiny> EXPLAIN (TYPE DISTRIBUTED) SELECT count(*) FROM nation;
                               Query Plan                               
------------------------------------------------------------------------
 Fragment 0 [SINGLE]                                                    
     Output layout: [count]                                             
     Output partitioning: SINGLE []                                     
     Output[columnNames = [_col0]]                                      
     │   Layout: [count:bigint]                                         
     │   Estimates: {rows: ? (?), cpu: ?, memory: ?, network: ?}        
     │   _col0 := count                                                 
     └─ Aggregate[type = FINAL]                                         
        │   Layout: [count:bigint]                                      
        │   Estimates: {rows: ? (?), cpu: ?, memory: ?, network: ?}     
        │   count := count("count_0")                                   
        └─ LocalExchange[partitioning = SINGLE]                         
           │   Layout: [count_0:bigint]                                 
           │   Estimates: {rows: ? (?), cpu: ?, memory: ?, network: ?}  
           └─ RemoteSource[sourceFragmentIds = [1]]                     
                  Layout: [count_0:bigint]                              
                  Estimates:                                            
                                                                        
 Fragment 1 [SOURCE]                                                    
     Output layout: [count_0]                                           
     Output partitioning: SINGLE []                                     
     Aggregate[type = PARTIAL]                                          
     │   Layout: [count_0:bigint]                                       
     │   Estimates:                                                     
     │   count_0 := count(*)                                            
     └─ TableScan[table = tpch:tiny:nation]                             
            Layout: []                                                  
            Estimates: {rows: 25 (0B), cpu: 0, memory: 0B, network: 0B} 
                                                                        
                                                                        
(1 row)

arhimondr · 2022-07-20T14:24:36Z

The Transition query to RUNNING state when at least one task is running commit is needed as previously the bug was not visible, as there was always a stage with an exactly a single task

losipiuk · 2022-07-20T14:29:59Z

core/trino-main/src/main/java/io/trino/execution/scheduler/SqlQueryScheduler.java

                        ImmutableMultimap.of());
                stageExecution.schedulingComplete();
                remoteTask.ifPresent(task -> coordinatorTaskManager.addSourceTaskFailureListener(task.getTaskId(), failureReporter));
+                if (queryStateMachine.getQueryState() == STARTING && remoteTask.isPresent()) {


why just STARTING?

The order of states is STARTING -> RUNNING -> FINSHING -> FINISHED | FAILED. If it's already in a later stage there's no need to transition

Technically you do not need if (queryStateMachine.getQueryState() == STARTING) at all as queryStateMachine.transitionToRunning does the check (it does queryState.setIf(RUNNING, currentState -> currentState.ordinal() < RUNNING.ordinal());)

losipiuk

second commit ok. I need some explanation on the first one.

losipiuk · 2022-07-20T15:25:50Z

core/trino-main/src/main/java/io/trino/execution/scheduler/SqlQueryScheduler.java

                        ScheduleResult result = stageSchedulers.get(stageExecution.getStageId())
                                .schedule();

+                        if (stateMachine.getState() == DistributedStagesSchedulerState.PLANNED && stageExecution.getAllTasks().size() > 0) {


same here: if (stateMachine.getState() == DistributedStagesSchedulerState.PLANNED is not needed

I wanted to avoid calling stageExecution.getAllTasks() every time as it has to acquire a list of tasks under a lock and made the similar code in the coordinator scheduler consistent with this. Likely it won't be a problem, but at the same time I wonder if it introduces enough confusion to justify extra CPU cycles?

I see. Keep it then

cla-bot bot added the cla-signed label Jul 18, 2022

github-actions bot added the tests:hive label Jul 18, 2022

arhimondr force-pushed the disable-force-single-node-output branch from b1d550f to 87056bb Compare July 18, 2022 23:16

arhimondr requested review from dain and losipiuk July 18, 2022 23:29

arhimondr added 2 commits July 18, 2022 22:14

Transition query to RUNNING state when at least one task is running

a42e287

Stage execution could remain in the same SCHEDULING state while already running some tasks and not transition the state further until more tasks are scheduled.

arhimondr force-pushed the disable-force-single-node-output branch from 87056bb to 32d1f29 Compare July 19, 2022 02:18

findepi requested a review from sopel39 July 19, 2022 11:02

arhimondr requested a review from findepi July 20, 2022 00:37

losipiuk reviewed Jul 20, 2022

View reviewed changes

losipiuk approved these changes Jul 20, 2022

View reviewed changes

arhimondr merged commit ea28b6a into trinodb:master Jul 20, 2022

arhimondr deleted the disable-force-single-node-output branch July 20, 2022 20:39

colebow mentioned this pull request Jul 20, 2022

Add Trino 391 release notes #13216

Merged

github-actions bot added this to the 391 milestone Jul 20, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Set optimizer.force-single-node-output to false by default#13217

Set optimizer.force-single-node-output to false by default#13217
arhimondr merged 2 commits intotrinodb:masterfrom
arhimondr:disable-force-single-node-output

arhimondr commented Jul 18, 2022

Uh oh!

sopel39 commented Jul 19, 2022 •

edited

Loading

Uh oh!

findepi commented Jul 19, 2022

Uh oh!

arhimondr commented Jul 19, 2022

Uh oh!

arhimondr commented Jul 20, 2022

Uh oh!

losipiuk Jul 20, 2022

Uh oh!

arhimondr Jul 20, 2022

Uh oh!

losipiuk Jul 20, 2022

Uh oh!

losipiuk left a comment

Uh oh!

losipiuk Jul 20, 2022

Uh oh!

arhimondr Jul 20, 2022

Uh oh!

losipiuk Jul 20, 2022

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

4 participants

Conversation

arhimondr commented Jul 18, 2022

Description

Related issues, pull requests, and links

Documentation

Release notes

Uh oh!

sopel39 commented Jul 19, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

findepi commented Jul 19, 2022

Uh oh!

arhimondr commented Jul 19, 2022

Uh oh!

arhimondr commented Jul 20, 2022

Uh oh!

losipiuk Jul 20, 2022

Choose a reason for hiding this comment

Uh oh!

arhimondr Jul 20, 2022

Choose a reason for hiding this comment

Uh oh!

losipiuk Jul 20, 2022

Choose a reason for hiding this comment

Uh oh!

losipiuk left a comment

Choose a reason for hiding this comment

Uh oh!

losipiuk Jul 20, 2022

Choose a reason for hiding this comment

Uh oh!

arhimondr Jul 20, 2022

Choose a reason for hiding this comment

Uh oh!

losipiuk Jul 20, 2022

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

4 participants

sopel39 commented Jul 19, 2022 •

edited

Loading