* GlueingPartitioningOperator: It continuously receives data, and outputs batches of partitioned RACs. It maintains a last-partitioning-boundary of the last-pushed-RAC, and attempts to glue it with the next RAC it receives, ensuring that partitions are handled correctly, even across multiple RACs. You can check GlueingPartitioningOperatorTest for some good examples of the "glueing" work.
* PartitionSortOperator: It sorts rows inside partitioned RACs, on the sort columns. The input RACs it receives are expected to be "complete / separate" partitions of data.
Follow up to #17214, adds implementations for substituteCombiningFactory so that more
datasketches aggs can match projections, along with some projections tests for datasketches.
* Logger: Log context of DruidExceptions.
There is often interesting and unique information available in the
"context" of a DruidException. This information is additive to both
the message and the cause, and was missed when we log. This patch adds
the DruidException context to log messages whenever stack traces are
enabled.
* Only log nonempty contexts.
The javadoc for SegmentDescriptor discusses differences between it and
SegmentId, but misses the most important difference: SegmentDescriptor
can have a narrower interval than the segment being referenced.
Refactors a few things.
- Adds SemanticUtils maps to columns.
- Add some addAll functions to reduce duplication, and for future reuse.
- Refactor VariantColumnAndIndexSupplier to only take a SmooshedFileMapper instead.
- Refactor LongColumnSerializerV2 to have separate functions for serializing a value and null.
changes:
adds ExpressionProcessing.allowVectorizeFallback() and ExpressionProcessingConfig.allowVectorizeFallback(), defaulting to false until few remaining bugs can be fixed (mostly complex types and some odd interactions with mixed types)
add cannotVectorizeUnlessFallback functions to make it easy to toggle the default of this config, and easy to know what to delete when we remove it in the future
This fixes a race where, if there is no output at all, setAllDoneIfPossible
could be called twice (once when the output partitions future resolves, and
once when the batcher finishes). If the calls happen in that order, it would
try to create nil output channels both times, resulting in a "Channel already set"
error.
Stages can be instructed to exit before they finish, especially when a
downstream stage includes a "LIMIT". This patch has improvements related
to early-exiting stages.
Bug fix:
- WorkerStageKernel: Don't allow fail() to set an exception if the stage is
already in a terminal state (FINISHED or FAILED). If fail() is called while
in a terminal state, log the exception, then throw it away. If it's a
cancellation exception, don't even log it. This fixes a bug where a stage
that exited early could transition to FINISHED and then to FAILED, causing
the overall query to fail.
Performance:
- DartWorkerManager previously sent stopWorker commands to workers
even when "interrupt" was false. Now it only sends those commands when
"interrupt" is true. The method javadoc already claimed this is what the
method did, but the implementation did not match the javadoc. This reduces
the number of RPCs by 1 per worker per query.
Quieter logging:
- In ReadableByteChunksFrameChannel, skip logging exception from setError if
the channel has been closed. Channels are closed when readers are done with
them, so at that point, we wouldn't be interested in the errors.
- In RunWorkOrder, skip calling notifyListener on failure of the main work,
in the case when stop() has already been called. The stop() method will
set its own error using CanceledFault. This enables callers to detect
when a stage was canceled vs. failed for some other reason.
- In WorkerStageKernel, skip logging cancellation errors in fail(). This is
made possible by the previous change in RunWorkOrder.
* adds support for `UNNEST` expressions
* introduces `LogicalUnnestRule` to transform a `Correlate` doing UNNEST into a `LogicalUnnest`
* `UnnestInputCleanupRule` could move the final unnested expr into the `LogicalUnnest` itself (usually its an `mv_to_array` expression)
* enhanced source unwrapping to utilize `FilteredDataSource` if it looks right
This patch adds a profile of MSQ named "Dart" that runs on Brokers and
Historicals, and which is compatible with the standard SQL query API.
For more high-level description, and notes on future work, refer to #17139.
This patch contains the following changes, grouped into packages.
Controller (org.apache.druid.msq.dart.controller):
The controller runs on Brokers. Main classes are,
- DartSqlResource, which serves /druid/v2/sql/dart/.
- DartSqlEngine and DartQueryMaker, the entry points from SQL that actually
run the MSQ controller code.
- DartControllerContext, which configures the MSQ controller.
- DartMessageRelays, which sets up relays (see "message relays" below) to read
messages from workers' DartControllerClients.
- DartTableInputSpecSlicer, which assigns work based on a TimelineServerView.
Worker (org.apache.druid.msq.dart.worker)
The worker runs on Historicals. Main classes are,
- DartWorkerResource, which supplies the regular MSQ WorkerResource, plus
Dart-specific APIs.
- DartWorkerRunner, which runs MSQ worker code.
- DartWorkerContext, which configures the MSQ worker.
- DartProcessingBuffersProvider, which provides processing buffers from
sliced-up merge buffers.
- DartDataSegmentProvider, which provides segments from the Historical's
local cache.
Message relays (org.apache.druid.messages):
To avoid the need for Historicals to contact Brokers during a query, which
would create opportunities for queries to get stuck, all connections are
opened from Broker to Historical. This is made possible by a message relay
system, where the relay server (worker) has an outbox of messages.
The relay client (controller) connects to the outbox and retrieves messages.
Code for this system lives in the "server" package to keep it separate from
the MSQ extension and make it easier to maintain. The worker-to-controller
ControllerClient is implemented using message relays.
Other changes:
- Controller: Added the method "hasWorker". Used by the ControllerMessageListener
to notify the appropriate controllers when a worker fails.
- WorkerResource: No longer tries to respond more than once in the
"httpGetChannelData" API. This comes up when a response due to resolved future
is ready at about the same time as a timeout occurs.
- MSQTaskQueryMaker: Refactor to separate out some useful functions for reuse
in DartQueryMaker.
- SqlEngine: Add "queryContext" to "resultTypeForSelect" and "resultTypeForInsert".
This allows the DartSqlEngine to modify result format based on whether a "fullReport"
context parameter is set.
- LimitedOutputStream: New utility class. Used when in "fullReport" mode.
- TimelineServerView: Add getDruidServerMetadata as a performance optimization.
- CliHistorical: Add SegmentWrangler, so it can query inline data, lookups, etc.
- ServiceLocation: Add "fromUri" method, relocating some code from ServiceClientImpl.
- FixedServiceLocator: New locator for a fixed set of service locations. Useful for
URI locations.
Fixes a mistake introduced in #16533 which can result in CursorGranularizer incorrectly trying to get values from a selector after calling cursor.advance because of a missing check for cursor.isDone
changes:
* filter index processing is now automatically ordered based on estimated 'cost', which is approximated based on how many expected bitmap operations are required to construct the bitmap used for the 'offset'
* cursorAutoArrangeFilters context flag now defaults to true, but can be set to false to disable cost based filter index sorting
Extracting a few miscellaneous non-functional changes from the batch supervisor branch:
- Replace anonymous inner classes with lambda expressions in the SQL supervisor manager layer
- Add explicit @Nullable annotations in DynamicConfigProviderUtils to make IDE happy
- Small variable renames (copy-paste error perhaps) and fix typos
- Add table name for this exception message: Delete the supervisor from the table[%s] in the database...
- Prefer CollectionUtils.isEmptyOrNull() over list == null || list.size() > 0. We can change the Precondition checks to throwing DruidException separately for a batch of APIs at a time.
* update docs for kafka lookup extension to specify correct extension ordering
* fix first line
* test with extension dependencies
* save work on dependency management
* working dependency graph
* working pull
* fix style
* fix style
* remove name
* load extension dependencies recursively
* generate depenencies on classloader creation
* add check for circular dependencies
* fix style
* revert style changes
* remove mutable class loader
* clean up class heirarchy
* extensions loader test working
* add unit tests
* pr comments
* fix unit tests
changes:
* add `ApplyFunction` support to vectorization fallback, allowing many of the remaining expressions to be vectorized
* add `CastToObjectVectorProcessor` so that vector engine can correctly cast any type
* add support for array and complex vector constants
* reduce number of cases which can block vectorization in expression planner to be unknown inputs (such as unknown multi-valuedness)
* fix array constructor expression, apply map expression to make actual evaluated type match the output type inference
* fix bug in array_contains where something like array_contains([null], 'hello') would return true if the array was a numeric array since the non-null string value would cast to a null numeric
* fix isNull/isNotNull to correctly handle any type of input argument
Text-based input formats like csv and tsv currently parse inputs only as strings, following the RFC4180Parser spec).
To workaround this, the web-console and other tools need to further inspect the sample data returned to sample data returned by the Druid sampler API to parse them as numbers.
This patch introduces a new optional config, tryParseNumbers, for the csv and tsv input formats. If enabled, any numbers present in the input will be parsed in the following manner -- long data type for integer types and double for floating-point numbers, and if parsing fails for whatever reason, the input is treated as a string. By default, this configuration is set to false, so numeric strings will be treated as strings.
* Use the whole frame when writing rows.
This patch makes the following adjustments to enable writing larger
single rows to frames:
1) RowBasedFrameWriter: Max out allocation size on the final doubling.
i.e., if the final allocation "naturally" would be 1 MiB but the
max frame size is 900 KiB, use 900 KiB rather than failing the 1 MiB
allocation.
2) AppendableMemory: In reserveAdditional, release the last block if it
is empty. This eliminates waste when a frame writer uses a
successive-doubling approach to find the right allocation size.
3) ArenaMemoryAllocator: Reclaim memory from the last allocation when
the last allocation is closed.
Prior to these changes, a single row could be much smaller than the
frame size and still fail to be added to the frame.
* Style.
* Fix test.
Currently, TaskDataSegmentProvider fetches the DataSegment from the Coordinator while loading the segment, but just discards it later. This PR refactors this to also return the DataSegment so that it can be used by workers without a separate fetch.
Previously, the processor used "remainingChannels" to track the number of
non-null entries of currentFrame. Now, "remainingChannels" tracks the
number of channels that are unfinished.
The difference is subtle. In the previous code, when an input channel
was blocked upon exiting nextFrame(), the "currentFrames" entry would be
null, and therefore the "remainingChannels" variable would be decremented.
After the next await and call to populateCurrentFramesAndTournamentTree(),
"remainingChannels" would be incremented if the channel had become
unblocked after awaiting.
This means that finished(), which returned true if remainingChannels was
zero, would not be reliable if called between nextFrame() and the
next await + populateCurrentFramesAndTournamentTree().
This patch changes things such that finished() is always reliable. This
fixes a regression introduced in PR #16911, which added a call to
finished() that was, at that time, unsafe.
Register a Ser-De for RowsAndColumns so that the window operator query running on leaf operators would be transferred properly on the wire. Would fix the empty response given by window queries without group by on the native engine.
* Speed up FrameFileTest, SuperSorterTest.
These are two heavily parameterized tests that, together, account for
about 60% of runtime in the test suite.
FrameFileTest changes:
1) Cache frame files in a static, rather than building the frame file
for each parameterization of the test.
2) Adjust TestArrayCursorFactory to cache the signature, rather than
re-creating it on each call to getColumnCapabilities.
SuperSorterTest changes:
1) Dramatically reduce the number of tests that run with
"maxRowsPerFrame" = 1. These are particularly slow due to writing so
many small files. Some still run, since it's useful to test edge cases,
but much fewer than before.
2) Reduce the "maxActiveProcessors" axis of the test from [1, 2, 4] to
[1, 3]. The aim is to reduce the number of cases while still getting
good coverage of the feature.
3) Reduce the "maxChannelsPerProcessor" axis of the test from [2, 3, 8]
to [2, 7]. The aim is to reduce the number of cases while still getting
good coverage of the feature.
4) Use in-memory input channels rather than file channels.
5) Defer formatting of assertion failure messages until they are needed.
6) Cache the cursor factory and its signature in a static.
7) Cache sorted test rows (used for verification) in a static.
* It helps to include the file.
* Style.
* abstract `IncrementalIndex` cursor stuff to prepare to allow for possibility of using different "views" of the data based on the cursor build spec
changes:
* introduce `IncrementalIndexRowSelector` interface to capture how `IncrementalIndexCursor` and `IncrementalIndexColumnSelectorFactory` read data
* `IncrementalIndex` implements `IncrementalIndexRowSelector`
* move `FactsHolder` interface to separate file
* other minor refactorings
* BaseWorkerClientImpl: Don't attempt to recover from a closed channel.
This patch introduces an exception type "ChannelClosedForWritesException",
which allows the BaseWorkerClientImpl to avoid retrying when the local
channel has been closed. This can happen in cases of cancellation.
* Add some test coverage.
* wip
* Add test coverage.
* Style.
* MSQ: Improved worker cancellation.
Four changes:
1) FrameProcessorExecutor now requires that cancellationIds be registered
with "registerCancellationId" prior to being used in "runFully" or "runAllFully".
2) FrameProcessorExecutor gains an "asExecutor" method, which allows that
executor to be used as an executor for future callbacks in such a way
that respects cancellationId.
3) RunWorkOrder gains a "stop" method, which cancels the current
cancellationId and closes the current FrameContext. It blocks until
both operations are complete.
4) Fixes a bug in RunAllFullyWidget where "processorManager.result()" was
called outside "runAllFullyLock", which could cause it to be called
out-of-order with "cleanup()" in case of cancellation or other error.
Together, these changes help ensure cancellation does not have races.
Once "cancel" is called for a given cancellationId, all existing processors
and running callbacks are canceled and exit in an orderly manner. Future
processors and callbacks with the same cancellationId are rejected
before being executed.
* Fix test.
* Use execute, which doesn't return, to avoid errorprone complaints.
* Fix some style stuff.
* Further enhancements.
* Fix style.
* MSQ: Rework memory management.
This patch reworks memory management to better support multi-threaded
workers running in shared JVMs. There are two main changes.
First, processing buffers and threads are moved from a per-JVM model to
a per-worker model. This enables queries to hold processing buffers
without blocking other concurrently-running queries. Changes:
- Introduce ProcessingBuffersSet and ProcessingBuffers to hold the
per-worker and per-work-order processing buffers (respectively). On Peons,
this is the JVM-wide processing pool. On Indexers, this is a per-worker
pool of on-heap buffers. (This change fixes a bug on Indexers where
excessive processing buffers could be used if MSQ tasks ran concurrently
with realtime tasks.)
- Add "bufferPool" argument to GroupingEngine#process so a per-worker pool
can be passed in.
- Add "druid.msq.task.memory.maxThreads" property, which controls the
maximum number of processing threads to use per task. This allows usage of
multiple processing buffers per task if admins desire.
- IndexerWorkerContext acquires processingBuffers when creating the FrameContext
for a work order, and releases them when closing the FrameContext.
- Add "usesProcessingBuffers()" to FrameProcessorFactory so workers know
how many sets of processing buffers are needed to run a given query.
Second, adjustments to how WorkerMemoryParameters slices up bundles, to
favor more memory for sorting and segment generation. Changes:
- Instead of using same-sized bundles for processing and for sorting,
workers now use minimally-sized processing bundles (just enough to read
inputs plus a little overhead). The rest is devoted to broadcast data
buffering, sorting, and segment-building.
- Segment-building is now limited to 1 concurrent segment per work order.
This allows each segment-building action to use more memory. Note that
segment-building is internally multi-threaded to a degree. (Build and
persist can run concurrently.)
- Simplify frame size calculations by removing the distinction between
"standard" and "large" frames. The new default frame size is the same
as the old "standard" frames, 1 MB. The original goal of of the large
frames was to reduce the number of temporary files during sorting, but
I think we can achieve the same thing by simply merging a larger number
of standard frames at once.
- Remove the small worker adjustment that was added in #14117 to account
for an extra frame involved in writing to durable storage. Instead,
account for the extra frame whenever we are actually using durable storage.
- Cap super-sorter parallelism using the number of output partitions, rather
than using a hard coded cap at 4. Note that in practice, so far, this cap
has not been relevant for tasks because they have only been using a single
processing thread anyway.
* Remove unused import.
* Fix errorprone annotation.
* Fixes for javadocs and inspections.
* Additional test coverage.
* Fix test.
changes:
* CursorHolder.isPreAggregated method indicates that a cursor has pre-aggregated data for all AggregatorFactory specified in a CursorBuildSpec. If true, engines should rewrite the query to use AggregatorFactory.getCombiningAggreggator, and column selector factories will provide selectors with the aggregator interediate type for the aggregator factory name
* Added groupby, timeseries, and topN support for CursorHolder.isPreAggregated
* Added synthetic test since no CursorHolder implementations support isPreAggregated at this point in time
If the GroupByMergingQueryRunner gets scheduled after the query timeout, it fails to clean up the processing tasks that have been scheduled. This can lead to unnecessary processing being done for the tasks whos results won't get consumed.
Akshat Jain
Hardik Bajaj
Clint Wylie
Gian Merlino
Shivam Garg
Karan Kumar
Adarsh Sanjeev
Zoltan Haindrich
Cece Mei
Abhishek Radhakrishnan
George Shiqi Wu
Adithya Chakilam
Sree Charan Manamala
Abhishek Agarwal
Pranav
Lasse Mammen
Laksh Singla