Use stats cache on error instead of the chained mechanism#15992
Use stats cache on error instead of the chained mechanism#15992gortiz merged 3 commits intoapache:masterfrom
Conversation
1e29882 to
16851af
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #15992 +/- ##
============================================
+ Coverage 62.90% 63.43% +0.53%
+ Complexity 1386 1357 -29
============================================
Files 2867 2912 +45
Lines 163354 166955 +3601
Branches 24952 25537 +585
============================================
+ Hits 102755 105916 +3161
- Misses 52847 53028 +181
- Partials 7752 8011 +259
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
| MultiStageQueryStats statsFromCancel = cancelWithStats(requestId, servers); | ||
| return result.withStats(statsFromCancel); | ||
| } | ||
| cancel(requestId, servers); |
There was a problem hiding this comment.
Why do we cancel here?
There was a problem hiding this comment.
This is useful to clean up the stats cache and solves the TODO below. I've just pushed a new version that is simpler to read (and only tries to cancel once).
This cancel method sends an asynchronous cancel request without waiting for a response, so it should be very inexpensive compared to all other GRPC messages we send (ie, rows sent using mailboxes or the plan itself).
|
Probably we should merge #16010 before this one. Otherwise we are going to receive too many useless log errors |
There was a problem hiding this comment.
Pull Request Overview
This PR updates the error-handling path to use the stats cache directly on failure rather than relying on the chained mailbox mechanism.
- Directly cancel and fetch stats via
cancelWithStatswhen a processing exception occurs. - Introduce
QueryResult.withStatsto merge new stats into an existing result. - Simplify cancellation logic in
submitAndReduce, removing the oldThrowablecatch and redundant cleanup.
Comments suppressed due to low confidence (3)
pinot-query-runtime/src/main/java/org/apache/pinot/query/service/dispatch/QueryDispatcher.java:171
- No existing tests cover the new error stats path when
result.getProcessingException()is non-null. Add a unit test to verify thatwithStatscorrectly incorporates stats on failure.
if (result.getProcessingException() != null) {
pinot-query-runtime/src/main/java/org/apache/pinot/query/service/dispatch/QueryDispatcher.java:791
- [nitpick] The new
withStatsmethod is public but lacks a Javadoc comment explaining its behavior and intended usage. Adding documentation will improve maintainability.
public QueryResult withStats(MultiStageQueryStats newQueryStats) {
pinot-query-runtime/src/main/java/org/apache/pinot/query/service/dispatch/QueryDispatcher.java:182
- The cleanup of
_serversByQuery(previously removed) is missing; without it, entries may accumulate when query cancellation is disabled. Consider restoring or centralizing the removal to avoid memory leaks.
if (!cancelled) {
| if (isQueryCancellationEnabled()) { | ||
| _serversByQuery.remove(requestId); | ||
| if (!cancelled) { | ||
| cancel(requestId, servers); |
There was a problem hiding this comment.
Sending one more request to each server on happy path does add overhead. Ideally we want to avoid this overhead. This is not a blocker, but we should try to figure out a way to minimize the overhead for happy path
…of the chained mechanism (apache#15992)" This reverts commit 3158efa
…ache#15992)" (apache#370) This reverts commit 3158efa.
…of the chained mechanism (apache#15992)" This reverts commit 3158efa
* [Query Resource Isolation] Workload Configs (apache#15109) * Workload Configs * workload config * Add API * config * Change config structure * Propagation strategy * Fix style check * Cost spliting on update * Table addition propagation * perf * Tests * test * test 2 * Review comments 1 * review comments 3 * review comments 3 * name change * review comments 4 * Fix TableDoesNotExistError for hybrid tables in MSE queries in controller API (apache#16102) * Make ThreadResourceUsageProvider a Helper/Utility Class. (apache#16051) * ThreadResourceUsageProvider is a helper class. ThreadResourceContext tracks resource usage. Fix updateConcurrently * Rename to ThreadResourceSnapshot * Clean up * Add javadoc * Done use auto closeable * Checkstyle * Fix compilation error * Add back removed functions in SPI * Remove private constructor because japicmp complains. * Add setThreadResourceUsageProvider because of backward-incompatible checks * Add setThreadResourceUsageProvider because of backward-incompatible checks * Fix test * Fix ThreadResourceSnapshot usage and tests * Store cpu sample in nanoseconds. * Reduce logs and improve logging when queries are terminated due to OOM. (apache#16172) * Dynamic PerQueryCPUMemAccountant Config on Servers (apache#16219) * Checkpoint * Register change handler * Fix bugs. Manually tested * Checkstyle * Tests * Add pre-check that values are default * Undo typo fix * Update QueryRunner to make use of window function overflow handling server configurations (apache#16108) * Add multistage thread limiting configs at the broker and server level (apache#16080) * Adding changes for supporting RLS (apache#16043) * Use stats cache on error instead of the chained mechanism (apache#15992) * Improve broker error messaging when broker is the one reporting the failure (apache#16076) * Introduce MSE active and passive timeouts (apache#16075) * Throttle SSE & MSE Tasks if Server heap usage is above a threshold (apache#16271) * Fix QueryScheduler constructor using class name. (apache#16280) * Fix QueryScheduler constructor using class name. * Fix test * [Query Resource Isolation] WorkloadBudgetManager and Host enforcement (apache#15798) * QRI - WorkloadBudgetManager implementation * Address review comments * Remove singleton & signature fix * Fix compatibility checker * Review comments * Move WorkloadBudgetManager to core. --------- Co-authored-by: praveenc7 <praveenkchaganlal@gmail.com> * Eliminate duplicate cancel attempts in PerQueryCPUMemAccountant (apache#16299) * Add basic 1 query tests * Add more tests * Add ability to remember cancel queries. * Clean up if conditions in killMostExpensiveQuery * Fix test failures. * Address review comments. * Use QueryCancelCallback to cancel queries from ThreadResourceUsageAccountant (apache#16142) * Remove all calls to System.gc() in PerQueryCPUMemAccountantFactory (apache#16374) * Initialize thread accountant just before serving queries (apache#16326) * Allow Reset of ThreadResourceUsageAccountant in Tracing.java (apache#16360) * Queries now self terminate if in panic mode. (apache#16380) * Queries now self terminate if in panic mode. * Add config test * Hard kill on critical level. * Fix configs * Separate anchor thread interruption. * Checkstyle * Review comments * remove code for critical level --------- Co-authored-by: Rajat Venkatesh <vrajat@users.noreply.github.com> * [Query Resource Isolation] Additonal Sampling for Broker and Server (apache#16164) * fix * sampling * Broker sampling * revert integ-test * Fix test failures * review comments * remove MSE * broker auth * remove per pruner & planner sample * Use Broker's accountant to sample in the request handler. (apache#16439) * [Query Resource Isolation] Workload Scheduler (apache#16018) * QRI - WorkloadBudgetManager implementation * Address review comments * scheduler * unit test * review comments: metrics, secondary, resource-manager * remove broker admission * Remove default budget --------- Co-authored-by: Vivek Iyer Vaidyanathan Iyer <vvaidyanathan@linkedin.com> * Cleanup deprecated methods in ThreadResourceUsageAccountant (apache#16479) * Remove unnecessary methods and config for ThreadResourceUsageAccountant (apache#16490) * Add tests for OOM Termination of MSE queries. (apache#16514) * Fix a flaky assert when testing OOM Cancellation of MSE Queries (apache#16533) * Disable Flaky Tests (apache#16554) This is a follow-up to apache#16533 The fix for a flaky test did not work. This PR disables these tests temporarily. * Use correlation ID instead of request id in PerQueryCpuMemAccountant (apache#16040) * [Query Resource Isolation]Interface for Workload Stats Collection (apache#16340) * Interface for Stats Collection * Additional comments * inherit * additional class comments * [Query Resource Isolation] Fix Refresh message (apache#16636) * Fix Refresh message * delete queryworkload message handler * info -> debug logs * reduce logging (apache#16698) * style check * [Query Workload Isolation] Cost-split support (apache#16672) * splits * Cost split * test * propagation entity change & java doc * Propagation scheme review comments * empty commit to trigger build * Reduce log for PerQueryCPUMemResourceUsageAccountant (apache#16642) --------- Co-authored-by: Rajat Venkatesh <1638298+vrajat@users.noreply.github.com> Co-authored-by: Yash Mayya <yash.mayya@gmail.com> Co-authored-by: Satwik Pachigolla <40644097+satwik-pachigolla@users.noreply.github.com> Co-authored-by: 9aman <35227405+9aman@users.noreply.github.com> Co-authored-by: Gonzalo Ortiz Jaureguizar <gortiz@users.noreply.github.com> Co-authored-by: Vivek Iyer Vaidyanathan <vvivekiyer@gmail.com> Co-authored-by: Xiaotian (Jackie) Jiang <17555551+Jackie-Jiang@users.noreply.github.com> Co-authored-by: Rajat Venkatesh <vrajat@users.noreply.github.com> Co-authored-by: Vivek Iyer Vaidyanathan Iyer <vvaidyanathan@linkedin.com>
* [Query Resource Isolation] Workload Configs (apache#15109) * Workload Configs * workload config * Add API * config * Change config structure * Propagation strategy * Fix style check * Cost spliting on update * Table addition propagation * perf * Tests * test * test 2 * Review comments 1 * review comments 3 * review comments 3 * name change * review comments 4 * Fix TableDoesNotExistError for hybrid tables in MSE queries in controller API (apache#16102) * Make ThreadResourceUsageProvider a Helper/Utility Class. (apache#16051) * ThreadResourceUsageProvider is a helper class. ThreadResourceContext tracks resource usage. Fix updateConcurrently * Rename to ThreadResourceSnapshot * Clean up * Add javadoc * Done use auto closeable * Checkstyle * Fix compilation error * Add back removed functions in SPI * Remove private constructor because japicmp complains. * Add setThreadResourceUsageProvider because of backward-incompatible checks * Add setThreadResourceUsageProvider because of backward-incompatible checks * Fix test * Fix ThreadResourceSnapshot usage and tests * Store cpu sample in nanoseconds. * Reduce logs and improve logging when queries are terminated due to OOM. (apache#16172) * Dynamic PerQueryCPUMemAccountant Config on Servers (apache#16219) * Checkpoint * Register change handler * Fix bugs. Manually tested * Checkstyle * Tests * Add pre-check that values are default * Undo typo fix * Update QueryRunner to make use of window function overflow handling server configurations (apache#16108) * Add multistage thread limiting configs at the broker and server level (apache#16080) * Adding changes for supporting RLS (apache#16043) * Use stats cache on error instead of the chained mechanism (apache#15992) * Improve broker error messaging when broker is the one reporting the failure (apache#16076) * Introduce MSE active and passive timeouts (apache#16075) * Throttle SSE & MSE Tasks if Server heap usage is above a threshold (apache#16271) * Fix QueryScheduler constructor using class name. (apache#16280) * Fix QueryScheduler constructor using class name. * Fix test * [Query Resource Isolation] WorkloadBudgetManager and Host enforcement (apache#15798) * QRI - WorkloadBudgetManager implementation * Address review comments * Remove singleton & signature fix * Fix compatibility checker * Review comments * Move WorkloadBudgetManager to core. --------- Co-authored-by: praveenc7 <praveenkchaganlal@gmail.com> * Eliminate duplicate cancel attempts in PerQueryCPUMemAccountant (apache#16299) * Add basic 1 query tests * Add more tests * Add ability to remember cancel queries. * Clean up if conditions in killMostExpensiveQuery * Fix test failures. * Address review comments. * Use QueryCancelCallback to cancel queries from ThreadResourceUsageAccountant (apache#16142) * Remove all calls to System.gc() in PerQueryCPUMemAccountantFactory (apache#16374) * Initialize thread accountant just before serving queries (apache#16326) * Allow Reset of ThreadResourceUsageAccountant in Tracing.java (apache#16360) * Queries now self terminate if in panic mode. (apache#16380) * Queries now self terminate if in panic mode. * Add config test * Hard kill on critical level. * Fix configs * Separate anchor thread interruption. * Checkstyle * Review comments * remove code for critical level --------- Co-authored-by: Rajat Venkatesh <vrajat@users.noreply.github.com> * [Query Resource Isolation] Additonal Sampling for Broker and Server (apache#16164) * fix * sampling * Broker sampling * revert integ-test * Fix test failures * review comments * remove MSE * broker auth * remove per pruner & planner sample * Use Broker's accountant to sample in the request handler. (apache#16439) * [Query Resource Isolation] Workload Scheduler (apache#16018) * QRI - WorkloadBudgetManager implementation * Address review comments * scheduler * unit test * review comments: metrics, secondary, resource-manager * remove broker admission * Remove default budget --------- Co-authored-by: Vivek Iyer Vaidyanathan Iyer <vvaidyanathan@linkedin.com> * Cleanup deprecated methods in ThreadResourceUsageAccountant (apache#16479) * Remove unnecessary methods and config for ThreadResourceUsageAccountant (apache#16490) * Add tests for OOM Termination of MSE queries. (apache#16514) * Fix a flaky assert when testing OOM Cancellation of MSE Queries (apache#16533) * Disable Flaky Tests (apache#16554) This is a follow-up to apache#16533 The fix for a flaky test did not work. This PR disables these tests temporarily. * Use correlation ID instead of request id in PerQueryCpuMemAccountant (apache#16040) * [Query Resource Isolation]Interface for Workload Stats Collection (apache#16340) * Interface for Stats Collection * Additional comments * inherit * additional class comments * [Query Resource Isolation] Fix Refresh message (apache#16636) * Fix Refresh message * delete queryworkload message handler * info -> debug logs * reduce logging (apache#16698) * style check * [Query Workload Isolation] Cost-split support (apache#16672) * splits * Cost split * test * propagation entity change & java doc * Propagation scheme review comments * empty commit to trigger build * Reduce log for PerQueryCPUMemResourceUsageAccountant (apache#16642) * [refactor] Switching to RoutingManager for broker request handlers (apache#16442) Co-authored-by: Shaurya Chaturvedi <shauryachats@uber.com> * Fix broker request id generator to avoid generating same id (apache#16661) * Introduce QueryExecutionContext to manage query life cycle (apache#16728) * Introduce QueryExecutionContext to manage query life cycle 2 (apache#16728) --------- Co-authored-by: Rajat Venkatesh <1638298+vrajat@users.noreply.github.com> Co-authored-by: Yash Mayya <yash.mayya@gmail.com> Co-authored-by: Satwik Pachigolla <40644097+satwik-pachigolla@users.noreply.github.com> Co-authored-by: 9aman <35227405+9aman@users.noreply.github.com> Co-authored-by: Gonzalo Ortiz Jaureguizar <gortiz@users.noreply.github.com> Co-authored-by: Vivek Iyer Vaidyanathan <vvivekiyer@gmail.com> Co-authored-by: Xiaotian (Jackie) Jiang <17555551+Jackie-Jiang@users.noreply.github.com> Co-authored-by: Rajat Venkatesh <vrajat@users.noreply.github.com> Co-authored-by: Vivek Iyer Vaidyanathan Iyer <vvaidyanathan@linkedin.com> Co-authored-by: Shaurya Chaturvedi <shauryachats@gmail.com> Co-authored-by: Shaurya Chaturvedi <shauryachats@uber.com>
This PR changes the mechanism used to obtain stats on error. Instead of using the normal mailbox mechanism, it uses the alternative mode used on timeout. This increases the precision in cases where more than one worker fails (ie an invalid cast) or in queries with more than a few stages.