Fixes inclusion of all stream partitions in all tasks.
The PR (Adds Idle feature to `SeekableStreamSupervisor` for inactive stream) - https://github.com/apache/druid/pull/13144 updates the resulting lag calculation map in `KafkaSupervisor` to include all the latest partitions from the stream to set the idle state accordingly rather than the previous way of lag calculation only for the partitions actively being read from the stream. This led to an explosion of metrics in lag reports in cases where 1000s of tasks per supervisor are present.
Changes:
- Add a new method to generate lags for only those partitions a single task is actively reading from while updating the Supervisor reports.
We added compression to the latest/first pair storage, but
the code change was forcing new things to be persisted
with the new format, meaning that any segment created with
the new code cannot be read by the old code. Instead, we
need to default to creating the old format and then remove that default in a future version.
* Add string comparison methods to StringUtils, fix dictionary comparisons.
There are various places in Druid code where we assume that String.compareTo
is consistent with Unicode code-point ordering. Sadly this is not the case.
To help deal with this, this patch introduces the following helpers:
1) compareUnicode: Compares two Strings in Unicode code-point order.
2) compareUtf8: Compares two UTF-8 byte arrays in Unicode code-point order.
Equivalent to comparison as unsigned bytes.
3) compareUtf8UsingJavaStringOrdering: Compares two UTF-8 byte arrays, or
ByteBuffers, in a manner consistent with String.compareTo.
There is no helper for comparing two Strings in a manner consistent
with String.compareTo, because for that we can use compareTo directly.
The patch also fixes an inconsistency between the String and UTF-8
dictionary GenericIndexed flavors of string-typed columns: they were
formerly using incompatible comparators.
* Adjust test.
* FrontCodedIndexed updates.
* Add test.
* Fix comments.
Segment assignments can take very long due to the strategy cost computation
for a large number of segments. This commit allows segment assignments to be
done in a round-robin fashion within a tier. Only segment balancing takes cost-based
decisions to move segments around.
Changes
- Add dynamic config `useRoundRobinSegmentAssignment` with default value false
- Add `RoundRobinServerSelector`. This does not implement the `BalancerStrategy`
as it does not conform to that contract and may also be used in conjunction with a
strategy (round-robin for `RunRules` and a cost strategy for `BalanceSegments`)
- Drops are still cost-based even when round-robin assignment is enabled.
Druid catalog basics
Catalog object model for tables, columns
Druid metadata DB storage (as an extension)
REST API to update the catalog (as an extension)
Integration tests
Model only: no planner integration yet
* SeekableStreamSupervisor: Don't enqueue duplicate notices.
Similar goal to #12018, but more aggressive. Don't enqueue a notice at
all if it is equal to one currently in the queue.
* Adjustments from review.
* Update indexing-service/src/test/java/org/apache/druid/indexing/overlord/supervisor/NoticesQueueTest.java
Co-authored-by: Kashif Faraz <kashif.faraz@gmail.com>
Co-authored-by: Kashif Faraz <kashif.faraz@gmail.com>
* Use standard library to correctly glob and stop at the correct folder structure when filtering cloud objects.
Removed:
import org.apache.commons.io.FilenameUtils;
Add:
import java.nio.file.FileSystems;
import java.nio.file.PathMatcher;
import java.nio.file.Paths;
* Forgot to update CloudObjectInputSource as well.
* Fix tests.
* Removed unused exceptions.
* Able to reduced user mistakes, by removing the protocol and the bucket on filter.
* add 1 more test.
* add comment on filterWithoutProtocolAndBucket
* Fix lint issue.
* Fix another lint issue.
* Replace all mention of filter -> objectGlob per convo here:
https://github.com/apache/druid/pull/13027#issuecomment-1266410707
* fix 1 bad constructor.
* Fix the documentation.
* Don’t do anything clever with the object path.
* Remove unused imports.
* Fix spelling error.
* Fix incorrect search and replace.
* Addressing Gian’s comment.
* add filename on .spelling
* Fix documentation.
* fix documentation again
Co-authored-by: Didip Kerabat <didip@apple.com>
* Fix typo
* Fix some spacing
* Add missing fields
* Cleanup table spacing
* Remove durable storage docs again
Thanks Brian for pointing out previous discussions.
* Update docs/multi-stage-query/reference.md
Co-authored-by: Charles Smith <techdocsmith@gmail.com>
* Mark codes as code
* And even more codes as code
* Another set of spaces
* Combine `ColumnTypeNotSupported`
Thanks Karan.
* More whitespaces and typos
* Add spelling and fix links
Co-authored-by: Charles Smith <techdocsmith@gmail.com>
* HttpPostEmitter back off send() busy-loop
The HttpPostEmitter gets in a loop until the flush timeout can be
triggered, OR until some new events arrive that reset the minimum
batch fill timeout delay. As a tactical fix, this introduces a
simple backoff delay to the send loop to prevent spamming logs.
* Update core/src/main/java/org/apache/druid/java/util/emitter/core/HttpPostEmitter.java
Co-authored-by: Frank Chen <frankchen@apache.org>
Co-authored-by: Frank Chen <frankchen@apache.org>
Changes:
- Add a metric for partition-wise kafka/kinesis lag for streaming ingestion.
- Emit lag metrics for streaming ingestion when supervisor is not suspended and state is in {RUNNING, IDLE, UNHEALTHY_TASKS, UNHEALTHY_SUPERVISOR}
- Document metrics
* scratch
* s3 ls fix, add docs
* add documentation, update method name
* Add tests, address commits, change default value of the helper
* fix test
* update the default value of config, remove initial delay config
* Trigger Build
* update class
* add more tests
* docs update
* spellcheck
* remove ioe from the signature
* add back dmmy constructor for initialization
* fix guice bindings, intellij inspections
`cachingCost` strategy has some discrepancies when compared to cost strategy.
This commit addresses two of these by retaining the same behaviour as the `cost` strategy
when computing the cost of moving a segment to a server:
- subtract the self cost of a segment if it is being served by the target server
- subtract the cost of segments that are marked to be dropped
Other changes:
- Add tests to verify fixed strategy. These tests would fail without the fixes made to `CachingCostStrategy.computeCost()`
- Fix the definition of the segment related metrics in the docs.
- Fix some docs issues introduced in #13181
* MSQ: Fix task lock checking during publish, fix lock priority.
Fixes two issues:
1) ControllerImpl did not properly check the return value of
SegmentTransactionalInsertAction when doing a REPLACE. This could cause
it to not realize that its locks were preempted.
2) Task lock priority was the default of 0. It should be the higher
batch default of 50. The low priority made it possible for MSQ tasks
to be preempted by compaction tasks, which is not desired.
* Restructuring, add docs.
* Add performSegmentPublish tests.
* Fix tests.
* Compaction: Fetch segments one at a time on main task; skip when possible.
Compact tasks include the ability to fetch existing segments and determine
reasonable defaults for granularitySpec, dimensionsSpec, and metricsSpec.
This is a useful feature that makes compact tasks work well even when the
user running the compaction does not have a clear idea of what they want
the compacted segments to be like.
However, this comes at a cost: it takes time, and disk space, to do all
of these fetches. This patch improves the situation in two ways:
1) When segments do need to be fetched, download them one at a time and
delete them when we're done. This still takes time, but minimizes the
required disk space.
2) Don't fetch segments on the main compact task when they aren't needed.
If the user provides a full granularitySpec, dimensionsSpec, and
metricsSpec, we can skip it.
* Adjustments.
* Changes from code review.
* Fix logic for determining rollup.
* MSQ: Consider PARTITION_STATS_MAX_BYTES in WorkerMemoryParameters.
This consideration is important, because otherwise we can run out of
memory due to large statistics-tracking objects.
* Improved calculations.
Clint Wylie
Tejaswini Bandlamudi
imply-cheddar
Gian Merlino
Kashif Faraz
dependabot[bot]
Laksh Singla
Adarsh Sanjeev
Paul Rogers
Jill Osborne
Didip Kerabat
Andreas Maechler
AmatyaAvadhanula
Jason Koch
Churro
Vadim Ogievetsky
Rohan Garg
Abhishek Agarwal