You are viewing a plain text version of this content. The canonical link for it is here.

Posted to dev@gobblin.apache.org by GitBox <gi...@apache.org> on 2020/10/16 05:45:16 UTC

[GitHub] [incubator-gobblin] autumnust opened a new pull request #3128: [GOBBLIN-1290] Auto-tune ORC writer parameters

autumnust opened a new pull request #3128:
URL: https://github.com/apache/incubator-gobblin/pull/3128

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](https://issues.apache.org/jira/browse/GOBBLIN/) issues and references them in the PR title. For example, "[GOBBLIN-XXX] My Gobblin PR"
- https://issues.apache.org/jira/browse/GOBBLIN-1290

### Description
- [ ] Here are some details about my PR, including screenshots (if applicable):
Here is a brief intro for the memory consumption within the streaming ingestor:

Physical Memory (Usually 4G)
- 50% for ORC writer (2G): This is not a hard cap, it is primarily used for scaling down the threshold for each writer registered in the memory manager. So it is possible to exceed this limit if the number of writers increases when, for example, it needs to deal with portions of late data.
During normal traffic (non catching up/backfilling mode), having 2~3 writers running in parallel is quite normal due to hourly-boundary interval and late data arrival.
- Metrics objects and other Gobblin overhead : Usually around ~400MB, Typical consumer:
- TaskExecutor
- Kafka Extractor
Non heap memory overhead (600MB, configurable)

Given the approximate data points above, we should have no more than ~500MB raw buffer space for each writer. Given the default number of rows between memory check (5000), that means record size should be smaller than 100K. The record size we measured in the streaming pipeline comes from serialized key size plus serialized value size, when they are stored in TreeWriter, the value could be a bit more larger given the object overhead. Therefore there's a scaling factor being used to indicate this.

The equation for the auto-tuning is:
#rowBetweenCheck = actualHeapSize * ratio (50%) / parallelsimOfWriters / (serializedRecordSize * scalingFactor)

Tested with several heavy topics internally and it works well.

Other than the major changes, also fix two small issues:
- build.gradle in gobblin-service module hits a problem and resolved by referring to: https://stackoverflow.com/questions/36079404/errorcannot-configure-the-publishing-extension-after-it-has-been-accessed
- update gitignore file

### Tests
- [ ] My PR adds the following unit tests __OR__ does not need testing for this extremely good reason:

### 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](http://chris.beams.io/posts/git-commit/)":
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"

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org