[GOBBLIN-1957] GobblinOrcwriter improvements for large records by Will-Lo · Pull Request #3828 · apache/gobblin

Will-Lo · 2023-11-15T21:05:57Z

Dear Gobblin maintainers,

Please accept this PR. I understand that it will not be reviewed until I have checked off all the steps below!

JIRA

My PR addresses the following Gobblin JIRA issues and references them in the PR title. For example, "[GOBBLIN-XXX] My Gobblin PR"
- https://issues.apache.org/jira/browse/GOBBLIN-1957

Description

Here are some details about my PR, including screenshots (if applicable):

GobblinORCWriter self tune uses a number of metrics to determine how large their buffers should be for both its own internal buffer used for conversion and the native ORC writer buffer. However when there are very large record sizes (100s of kb) the buffers default max size (e.g. 1000) can still hold a very large amount of data. Observed performance would be hundreds of megabytes to even a gigabyte depending on the configured batch size maximums.

We want a configuration to impose a maximum buffer max size so that large records in the buffer do not exceed the size of a stripe, so when it is added to the native ORC Writer, the native orc writer should be flushing its records and freeing the memory.

This PR adds the configuration orcWriter.auto.selfTune.max.buffer.orc.stripe which when enabled to true, will restrict the max buffer size to be the value of the ORC stripe size divided by the estimated record size recorded. Since it can still tune the buffer sizes in between the min value (1) and this maximum, it will not overfill the buffers like previously recorded.

Since this configuration may impact performance, hid it under a feature flag as it will lead to less OOM but more frequent memory checks may impact write speed. However since it should generally be checking as often as it has enough memory to fill one stripe, this should be acceptable.

Tests

My PR adds the following unit tests OR does not need testing for this extremely good reason:

Unit tests, tested on large topic, reduction in OOM observed

Commits

My commits all reference JIRA issues in their subject lines, and I have squashed multiple commits if they address the same issue. In addition, my commits follow the guidelines from "How to write a good git commit message":
1. Subject is separated from body by a blank line
2. Subject is limited to 50 characters
3. Subject does not end with a period
4. Subject uses the imperative mood ("add", not "adding")
5. Body wraps at 72 characters
6. Body explains "what" and "why", not "how"

codecov-commenter · 2023-11-15T21:22:12Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 47.54%. Comparing base (dd17bed) to head (7c4bc0a).
⚠️ Report is 309 commits behind head on master.

Additional details and impacted files

@@             Coverage Diff              @@
##             master    #3828      +/-   ##
============================================
+ Coverage     47.53%   47.54%   +0.01%     
- Complexity    11035    11038       +3     
============================================
  Files          2156     2156              
  Lines         85377    85398      +21     
  Branches       9491     9493       +2     
============================================
+ Hits          40581    40601      +20     
  Misses        41099    41099              
- Partials       3697     3698       +1

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

ZihanLi58 · 2023-11-15T21:49:16Z

gobblin-modules/gobblin-orc/src/main/java/org/apache/gobblin/writer/GobblinBaseOrcWriter.java

    this.typeDescription = getOrcSchema();
    this.selfTuningWriter = properties.getPropAsBoolean(GobblinOrcWriterConfigs.ORC_WRITER_AUTO_SELFTUNE_ENABLED, false);
    this.validateORCAfterClose = properties.getPropAsBoolean(GobblinOrcWriterConfigs.ORC_WRITER_VALIDATE_FILE_AFTER_CLOSE, false);
+    this.rowCheckFactor = properties.getPropAsInt(GobblinOrcWriterConfigs.ORC_WRITER_BATCHSIZE_ROWCHECK_FACTOR, GobblinOrcWriterConfigs.DEFAULT_ORC_WRITER_BATCHSIZE_ROWCHECK_FACTOR);


why do we make this one configurable?

I think the default of 5 which I made static was pretty aggressive, so wanted to make it tunable to measure performance

ZihanLi58 · 2023-11-15T21:50:02Z

gobblin-modules/gobblin-orc/src/main/java/org/apache/gobblin/writer/GobblinBaseOrcWriter.java

        GobblinOrcWriterConfigs.DEFAULT_MIN_ORC_WRITER_ROWCHECK);
    this.orcFileWriterMaxRowsBetweenCheck = properties.getPropAsInt(GobblinOrcWriterConfigs.ORC_WRITER_MAX_ROWCHECK,
        GobblinOrcWriterConfigs.DEFAULT_MAX_ORC_WRITER_ROWCHECK);
+    this.enableLimitBufferSizeOrcStripe = properties.getPropAsBoolean(GobblinOrcWriterConfigs.ORC_WRITER_ENABLE_BUFFER_LIMIT_ORC_STRIPE, false);


Same for this, do you plan to clean up the code afterwards?

Yeah didn't want to introduce an unwanted regression that would be hard to roll back, can make this the static/default behavior if performance is good

homatthew · 2023-11-17T01:06:34Z

gobblin-modules/gobblin-orc/src/main/java/org/apache/gobblin/writer/GobblinBaseOrcWriter.java

    this.typeDescription = getOrcSchema();
    this.selfTuningWriter = properties.getPropAsBoolean(GobblinOrcWriterConfigs.ORC_WRITER_AUTO_SELFTUNE_ENABLED, false);
    this.validateORCAfterClose = properties.getPropAsBoolean(GobblinOrcWriterConfigs.ORC_WRITER_VALIDATE_FILE_AFTER_CLOSE, false);
+    this.rowCheckFactor = properties.getPropAsInt(GobblinOrcWriterConfigs.ORC_WRITER_BATCHSIZE_ROWCHECK_FACTOR, GobblinOrcWriterConfigs.DEFAULT_ORC_WRITER_BATCHSIZE_ROWCHECK_FACTOR);


I find this variable name a little ambiguous. The longer form of batchSizeRowCheckFactor makes a little more sense to me

ZihanLi58

+1

* [GOBBLIN-1921] Properly handle reminder events (apache#3790) * Add millisecond level precision to timestamp cols & proper timezone conversion - existing tests pass with minor modifications * Handle reminder events properly * Fix compilation errors & add isReminder flag * Add unit tests * Address review comments * Add newline to address comment * Include reminder/original tag in logging * Clarify timezone issues in comment --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1924] Reminder event flag true (apache#3795) * Set reminder event flag to true for reminders * Update unit tests * remove unused variable --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1923] Add retention for lease arbiter table (apache#3792) * Add retention for lease arbiter table * Replace blocking thread with scheduled thread pool executor * Make Calendar instance thread-safe * Rename variables, make values more clear * Update timestamp related cols --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * Change debug statements to info temporarily to debug (apache#3796) Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1926] Fix Reminder Event Epsilon Comparison (apache#3797) * Fix Reminder Event Epsilon Comparison * Add TODO comment --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1930] Improve Multi-active related logs and metrics (apache#3800) * Improve Multi-active related logs and metrics * Add more metrics and logs around forwarding dag action to DagManager * Improve logs in response to review comments * Replace flow execution id with trigger timestamp from multi-active * Update flow action execution id within lease arbiter * Fix test & make Lease Statuses more lean * Update javadoc --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * add dataset root in some common type of datasets * [GOBBLIN-1927] Add topic validation support in KafkaSource, and add TopicNameValidator (apache#3793) * * Add generic topic validation support * Add the first validator TopicNameValidator into the validator chain, as a refactor of existing codes * Refine to address comments * Refine --------- Co-authored-by: Tao Qin <tqin@linkedin.com> * [GOBBLIN-1931] Refactor dag action updating method & add clarifying comment (apache#3801) * Refactor dag action updating method & add clarifying comment * Log filtered out duplicate messages * logs and metrics for missing messages from change monitor * Only add gobblin.service prefix for dagActionStoreChangeMonitor --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1934] Monitor High Level Consumer queue size (apache#3805) * Emit metrics to monitor high level consumer queue size * Empty commit to trigger tests * Use BlockingQueue.size() func instead of atomic integer array * Remove unused import & add DagActionChangeMonitor prefix to metric * Refactor to avoid repeating code * Make protected variables private where possible * Fix white space --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1935] Skip null dag action types unable to be processed (apache#3807) * Skip over null dag actions from malformed messages * Add new metric for skipped messages --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1922]Add function in Kafka Source to recompute workUnits for filtered partitions (apache#3798) * add function in Kafka Source to recompute workUnits for filtered partitions * address comments * set default min container value to 1 * add condition when create empty wu * update the condition * Expose functions to fetch record partitionColumn value (apache#3810) * [GOBBLIN-1938] preserve x bit in manifest file based copy (apache#3804) * preserve x bit in manifest file based copy * fix project structure preventing running unit tests from intellij * fix unit test * [GOBBLIN-1919] Simplify a few elements of MR-related job exec before reusing code in Temporal-based execution (apache#3784) * Simplify a few elements of MR-related job exec before reusing code in Temporal-based execution * Add JSON-ification to several foundational config-state representations, plus encapsulated convience method `JobState.getJobIdFromProps` * Update javadoc comments * Encapsulate check for whether a path has the extension of a multi-work-unit * [GOBBLIN-1939] Bump AWS version to use a compatible version of Jackson with Gobblin (apache#3809) * Bump AWS version to use a compatible version of jackson with Gobblin * use shared aws version * [GOBBLIN-1937] Quantify Missed Work Completed by Reminders (apache#3808) * Quantify Missed Work Completed by Reminders Also fix bug to filter out heartbeat events before extracting field * Refactor changeMonitorUtils & add delimiter to metrics prefix * Re-order params to group similar ones --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * GOBBLIN-1933]Change the logic in completeness verifier to support multi reference tier (apache#3806) * address comments * use connectionmanager when httpclient is not cloesable * [GOBBLIN-1933] Change the logic in completeness verifier to support multi reference tier * add uite test * fix typo * change the javadoc * change the javadoc --------- Co-authored-by: Zihan Li <zihli@zihli-mn2.linkedin.biz> * [GOBBLIN-1943] Use AWS version 1.12.261 to fix a security vulnerability in the previous version (apache#3813) * [GOBBLIN-1941] Develop Temporal abstractions, including `Workload` for workflows of unbounded size through sub-workflow nesting (apache#3811) * Define `Workload` abstraction for Temporal workflows of unbounded size through sub-workflow nesting * Adjust Gobblin-Temporal configurability for consistency and abstraction * Define `WorkerConfig`, to pass the `TemporalWorker`'s configuration to the workflows and activities it hosts * Improve javadoc * Javadoc fixup * Minor changes * Update per review suggestions * Insert pause, to spread the load on the temporal server, before launch of each child workflow that may have direct leaves of its own * Appease findbugs by having `SeqSliceBackedWorkSpan::next` throw `NoSuchElementException` * Add comment * [GOBBLIN-1944] Add gobblin-temporal load generator for a single subsuming super-workflow with a configurable number of activities nested beneath (apache#3815) * Add gobblin-temporal load generator for a single subsuming super-workflow with a configurable number of activities nested beneath * Update per findbugs advice * Improve processing of int props * [GOBBLIN-1945] Implement Distributed Data Movement (DDM) Gobblin-on-Temporal `WorkUnit` evaluation (apache#3816) * Implement Distributed Data Movement (DDM) Gobblin-on-Temporal `WorkUnit` evaluation * Adjust work unit processing tuning for start-to-close timeout and nested execution branching * Rework `ProcessWorkUnitImpl` and fix `FileSystem` misuse; plus convenience abstractions to load `FileSystem`, `JobState`, and `StateStore<TaskState>` * Fix `FileSystem` resource lifecycle, uniquely name each workflow, and drastically reduce worker concurrent task execution * Heed findbugs advice * prep before commit * Improve processing of required props * Update comment in response to PR feedback * [GOBBLIN-1942] Create MySQL util class for re-usable methods and setup MysqlDagActio… (apache#3812) * Create MySQL util class for re-usable methods and setup MysqlDagActionStore retention * Add a java doc * Address review comments * Close scheduled executors on shutdown & clarify naming and comments * Remove extra period making config key invalid * implement Closeable * Use try with resources --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [Hotfix][GOBBLIN-1949] add option to detect malformed orc during commit (apache#3818) * add option to detect malformed ORC during commit phase * better logging * address comment * catch more generic exception * validate ORC file after close * move validate in between close and commit * syntax * whitespace * update log * [GOBBLIN-1948] Use same flowExecutionId across participants (apache#3819) * Use same flowExecutionId across participants * Set config field as well in new FlowSpec * Use gobblin util to create config * Rename function and move to util --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * Allow extension of functions in GobblinMCEPublisher and customization of fileList file metrics are calculated for (apache#3820) * [GOBBLIN-1951] Emit GTE when deleting corrupted ORC files (apache#3821) * [GOBBLIN-1951] Emit GTE when deleting corrupted ORC files This commit adds ORC file validation during the commit phase and deletes corrupted files. It also includes a test for ORC file validation. * Linter fixes * Add framework and unit tests for DagActionStoreChangeMonitor (apache#3817) * Add framework and unit tests for DagActionStoreChangeMonitor * Add more test cases and validation * Add header for new file * Move FlowSpec static function to Utils class * Remove unused import * Fix compile error * Fix unit tests --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1952] Make jobname shortening in GaaS more aggressive (apache#3822) * Make jobname shortening in GaaS more aggressive * Change long name prefix to flowgroup * Make KafkaTopicGroupingWorkUnitPacker pack with desired num of container (apache#3814) * Make KafkaTopicGroupingWorkUnitPacker pack with desired num of container * update comment * [GOBBLIN-1953] Add an exception message to orc writer validation GTE (apache#3826) * Fix FlowSpec Updating Function (apache#3823) * Fix FlowSpec Updating Function * makes Config object with FlowSpec mutable * adds unit test to ensure flow compiles after updating FlowSpec * ensure DagManager resilient to exceptions on leadership change * Only update Properties obj not Config to avoid GC overhead * Address findbugs error * Avoid updating or creating new FlowSpec objects by passing flowExecutionId directly to metadata * Remove changes that are not needed anymore * Add TODO to handle failed DagManager leadership change * Overload function and add more documentation --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * Emit metric to tune LeaseArbiter Linger metric (apache#3824) * Monitor number of failed persisting leases to tune linger * Increase default linger and epsilon values * Add metric for lease persisting success * Rename metrics --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1956]Make Kafka streaming pipeline be able to config the max poll records during runtime (apache#3827) * address comments * use connectionmanager when httpclient is not cloesable * add uite test * fix typo * [GOBBLIN-1956] Make Kafka streaming pipeline be able to config the max poll records during runtime * small refractor --------- Co-authored-by: Zihan Li <zihli@zihli-mn2.linkedin.biz> * Add semantics for failure on partial success (apache#3831) * Consistly handle Rest.li /flowexecutions KILL and RESUME actions (apache#3830) * [GOBBLIN-1957] GobblinOrcwriter improvements for large records (apache#3828) * WIP * Optimization to limit batchsize based on large record sizes * Address review * Use DB-qualified table ID as `IcebergTable` dataset descriptor (apache#3834) * [GOBBLIN-1961] Allow `IcebergDatasetFinder` to use separate names for source vs. destination-side DB and table (apache#3835) * Allow `IcebergDatasetFinder` to use separate names for source vs. destination-side DB and table * Adjust Mockito.verify to pass test * Prevent NPE in `FlowCompilationValidationHelper.validateAndHandleConcurrentExecution` (apache#3836) * Prevent NPE in `FlowCompilationValidationHelper.validateAndHandleConcurrentExecution` * improved `MultiHopFlowCompiler` javadoc * Delete Launch Action Events After Processing (apache#3837) * Delete launch action event after persisting * Fix default value for flowExecutionId retrieval from metadata map * Address review comments and add unit test * Code clean up --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1960] Emit audit count after commit in IcebergMetadataWriter (apache#3833) * Emit audit count after commit in IcebergMetadataWriter * Unit tests by extracting to a post commit * Emit audit count first * find bugs complaint * [GOBBLIN-1967] Add external data node for generic ingress/egress on GaaS (apache#3838) * Add external data node for generic ingress/egress on GaaS * Address reviews and cleanup * Use URI representation for external dataset descriptor node * Fix error message in containing check * Address review * [GOBBLIN-1971] Allow `IcebergCatalog` to specify the `DatasetDescriptor` name for the `IcebergTable`s it creates (apache#3842) * Allow `IcebergCatalog` to specify the `DatasetDescriptor` name for the `IcebergTable`s it creates * small method javadoc * [GOBBLIN-1970] Consolidate processing dag actions to one code path (apache#3841) * Consolidate processing dag actions to one code path * Delete dag action in failure cases too * Distinguish metrics for startup * Refactor to avoid duplicated code and create static metrics proxy class * Remove DagManager checks that don't apply on startup * Add test to check kill/resume dag action removal after processing * Remove unused import * Initialize metrics proxy with Null Pattern --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1972] Fix `CopyDataPublisher` to avoid committing post-publish WUs before they've actually run (apache#3844) * Fix `CopyDataPublisher` to avoid committing post-publish WUs before they've actually run * fixup findbugsMain * [GOBBLIN-1975] Keep job.name configuration immutable if specified on GaaS (apache#3847) * Revert "[GOBBLIN-1952] Make jobname shortening in GaaS more aggressive (apache#3822)" This reverts commit 5619a0a. * use configuration to keep specified jobname if enabled * Cleanup * [GOBBLIN-1974] Ensure Adhoc Flows can be Executed in Multi-active Scheduler state (apache#3846) * Ensure Adhoc Flows can be Executed in Multi-active Scheduler state * Only delete spec for adhoc flows & always after orchestration * Delete adhoc flows when dagManager is not present as well * Fix flaky test for scheduler * Add clarifying comment about failure recovery * Re-ordered private method * Move private methods again * Enforce sequential ordering of unit tests to make more reliable --------- Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> * [GOBBLIN-1973] Change Manifest distcp logic to compare permissions of source and dest files even when source is older (apache#3845) * change should copy logic * Add tests, address review * Fix checkstyle * Remove unused imports * [GOBBLIN-1976] Allow an `IcebergCatalog` to override the `DatasetDescriptor` platform name for the `IcebergTable`s it creates (apache#3848) * Allow an `IcebergCatalog` to override the `DatasetDescriptor` platform for the `IcebergTable`s it creates * fixup javadoc * Log when `PasswordManager` fails to load any master password (apache#3849) * [GOBBLIN-1968] Temporal commit step integration (apache#3829) Add commit step to Gobblin temporal workflow for job publish * Add codeql analysis * Make gradle specific * Add codeql as part of build script * Initialize codeql * Use separate workflow for codeql instead with custom build function as autobuild seems to not work * Add jdk jar for global dependencies script --------- Co-authored-by: umustafi <umust77@gmail.com> Co-authored-by: Urmi Mustafi <umustafi@linkedin.com> Co-authored-by: Arjun <abora@linkedin.com> Co-authored-by: Tao Qin <35046097+wsarecv@users.noreply.github.com> Co-authored-by: Tao Qin <tqin@linkedin.com> Co-authored-by: Hanghang Nate Liu <nate.hanghang.liu@gmail.com> Co-authored-by: Andy Jiang <andy.jiang99@outlook.com> Co-authored-by: Kip Kohn <ckohn@linkedin.com> Co-authored-by: Zihan Li <zihli@linkedin.com> Co-authored-by: Zihan Li <zihli@zihli-mn2.linkedin.biz> Co-authored-by: Matthew Ho <mho@linkedin.com>

Will-Lo added 2 commits November 15, 2023 16:10

WIP

5e777af

Optimization to limit batchsize based on large record sizes

b856009

Will-Lo force-pushed the orcwriter-improvements-large-records branch from ab3893d to b856009 Compare November 15, 2023 21:10

ZihanLi58 reviewed Nov 15, 2023

View reviewed changes

homatthew reviewed Nov 17, 2023

View reviewed changes

Address review

7c4bc0a

ZihanLi58 approved these changes Nov 20, 2023

View reviewed changes

Will-Lo merged commit 8b89075 into apache:master Nov 21, 2023

Will-Lo deleted the orcwriter-improvements-large-records branch November 21, 2023 16:17

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[GOBBLIN-1957] GobblinOrcwriter improvements for large records#3828

[GOBBLIN-1957] GobblinOrcwriter improvements for large records#3828
Will-Lo merged 3 commits intoapache:masterfrom
Will-Lo:orcwriter-improvements-large-records

Will-Lo commented Nov 15, 2023 •

edited

Loading

Uh oh!

codecov-commenter commented Nov 15, 2023 •

edited

Loading

Uh oh!

ZihanLi58 Nov 15, 2023

Uh oh!

Will-Lo Nov 16, 2023

Uh oh!

ZihanLi58 Nov 15, 2023

Uh oh!

Will-Lo Nov 16, 2023

Uh oh!

homatthew Nov 17, 2023

Uh oh!

ZihanLi58 left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

Will-Lo commented Nov 15, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

JIRA

Description

Tests

Commits

Uh oh!

codecov-commenter commented Nov 15, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

ZihanLi58 Nov 15, 2023

Choose a reason for hiding this comment

Uh oh!

Will-Lo Nov 16, 2023

Choose a reason for hiding this comment

Uh oh!

ZihanLi58 Nov 15, 2023

Choose a reason for hiding this comment

Uh oh!

Will-Lo Nov 16, 2023

Choose a reason for hiding this comment

Uh oh!

homatthew Nov 17, 2023

Choose a reason for hiding this comment

Uh oh!

ZihanLi58 left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Will-Lo commented Nov 15, 2023 •

edited

Loading

codecov-commenter commented Nov 15, 2023 •

edited

Loading