* MSQ: Fix validation of time position in collations.
It is possible for the collation to refer to a field that isn't mapped,
such as when the DML includes "CLUSTERED BY some_function(some_field)".
In this case, the collation refers to a projected column that is not
part of the field mappings. Prior to this patch, that would lead to an
out of bounds list access on fieldMappings.
This patch fixes the problem by identifying the position of __time in
the fieldMappings first, rather than retrieving each collation field
from fieldMappings.
Fixes a bug introduced in #16849.
* Fix test. Better warning message.
* Place __time in signatures according to sort order.
Updates a variety of places to put __time in row signatures according
to its position in the sort order, rather than always first, including:
- InputSourceSampler.
- ScanQueryEngine (in the default signature when "columns" is empty).
- Various StorageAdapters, which also have the effect of reordering
the column order in segmentMetadata queries, and therefore in SQL
schemas as well.
Follow-up to #16849.
* Fix compilation.
* Additional fixes.
* Fix.
* Fix style.
* Omit nonexistent columns from the row signature.
* Fix tests.
* Segments primarily sorted by non-time columns.
Currently, segments are always sorted by __time, followed by the sort
order provided by the user via dimensionsSpec or CLUSTERED BY. Sorting
by __time enables efficient execution of queries involving time-ordering
or granularity. Time-ordering is a simple matter of reading the rows in
stored order, and granular cursors can be generated in streaming fashion.
However, for various workloads, it's better for storage footprint and
query performance to sort by arbitrary orders that do not start with __time.
With this patch, users can sort segments by such orders.
For spec-based ingestion, users add "useExplicitSegmentSortOrder: true" to
dimensionsSpec. The "dimensions" list determines the sort order. To
define a sort order that includes "__time", users explicitly
include a dimension named "__time".
For SQL-based ingestion, users set the context parameter
"useExplicitSegmentSortOrder: true". The CLUSTERED BY clause is then
used as the explicit segment sort order.
In both cases, when the new "useExplicitSegmentSortOrder" parameter is
false (the default), __time is implicitly prepended to the sort order,
as it always was prior to this patch.
The new parameter is experimental for two main reasons. First, such
segments can cause errors when loaded by older servers, due to violating
their expectations that timestamps are always monotonically increasing.
Second, even on newer servers, not all queries can run on non-time-sorted
segments. Scan queries involving time-ordering and any query involving
granularity will not run. (To partially mitigate this, a currently-undocumented
SQL feature "sqlUseGranularity" is provided. When set to false the SQL planner
avoids using "granularity".)
Changes on the write path:
1) DimensionsSpec can now optionally contain a __time dimension, which
controls the placement of __time in the sort order. If not present,
__time is considered to be first in the sort order, as it has always
been.
2) IncrementalIndex and IndexMerger are updated to sort facts more
flexibly; not always by time first.
3) Metadata (stored in metadata.drd) gains a "sortOrder" field.
4) MSQ can generate range-based shard specs even when not all columns are
singly-valued strings. It merely stops accepting new clustering key
fields when it encounters the first one that isn't a singly-valued
string. This is useful because it enables range shard specs on
"someDim" to be created for clauses like "CLUSTERED BY someDim, __time".
Changes on the read path:
1) Add StorageAdapter#getSortOrder so query engines can tell how a
segment is sorted.
2) Update QueryableIndexStorageAdapter, IncrementalIndexStorageAdapter,
and VectorCursorGranularizer to throw errors when using granularities
on non-time-ordered segments.
3) Update ScanQueryEngine to throw an error when using the time-ordering
"order" parameter on non-time-ordered segments.
4) Update TimeBoundaryQueryRunnerFactory to perform a segment scan when
running on a non-time-ordered segment.
5) Add "sqlUseGranularity" context parameter that causes the SQL planner
to avoid using granularities other than ALL.
Other changes:
1) Rename DimensionsSpec "hasCustomDimensions" to "hasFixedDimensions"
and change the meaning subtly: it now returns true if the DimensionsSpec
represents an unchanging list of dimensions, or false if there is
some discovery happening. This is what call sites had expected anyway.
* Fixups from CI.
* Fixes.
* Fix missing arg.
* Additional changes.
* Fix logic.
* Fixes.
* Fix test.
* Adjust test.
* Remove throws.
* Fix styles.
* Fix javadocs.
* Cleanup.
* Smoother handling of null ordering.
* Fix tests.
* Missed a spot on the merge.
* Fixups.
* Avoid needless Filters.and.
* Add timeBoundaryInspector to test.
* Fix tests.
* Fix FrameStorageAdapterTest.
* Fix various tests.
* Use forceSegmentSortByTime instead of useExplicitSegmentSortOrder.
* Pom fix.
* Fix doc.
This PR generally improves the working of WriteOutBytes and WriteOutMedium. Some analysis of usage of TmpFileSegmentWriteOutMedium shows that they periodically get used for very small things. The overhead of creating a tmp file is actually very large. To improve the performance in these cases, this PR modifies TmpFileSegmentWriteOutMedium to return a heap-based WriteOutBytes that falls back to making a tmp file when it actually fills up.
---------
Co-authored-by: imply-cheddar <eric.tschetter@imply.io>
* Add type coercion and null check to left, right, repeat exprs.
These exprs shouldn't validate types; they should coerce types. Coercion
is typical behavior for functions because it enables schema evolution.
The functions are also modified to check isNumericNull on the right-hand
argument. This was missing previously, which would erroneously cause
nulls to be treated as zeroes.
* Fix tests.
The specific error on a truncated file can vary based on how the final
frame of the truncated file is written. This patch loosens the check so
it passes regardless of how the truncated file is written.
changes:
* Added `CursorBuildSpec` which captures all of the 'interesting' stuff that goes into producing a cursor as a replacement for the method arguments of `CursorFactory.canVectorize`, `CursorFactory.makeCursor`, and `CursorFactory.makeVectorCursor`
* added new interface `CursorHolder` and new interface `CursorHolderFactory` as a replacement for `CursorFactory`, with method `makeCursorHolder`, which takes a `CursorBuildSpec` as an argument and replaces `CursorFactory.canVectorize`, `CursorFactory.makeCursor`, and `CursorFactory.makeVectorCursor`
* `CursorFactory.makeCursors` previously returned a `Sequence<Cursor>` corresponding to the query granularity buckets, with a separate `Cursor` per bucket. `CursorHolder.asCursor` instead returns a single `Cursor` (equivalent to 'ALL' granularity), and a new `CursorGranularizer` has been added for query engines to iterate over the cursor and divide into granularity buckets. This makes the non-vectorized engine behave the same way as the vectorized query engine (with its `VectorCursorGranularizer`), and simplifies a lot of stuff that has to read segments particularly if it does not care about bucketing the results into granularities.
* Deprecated `CursorFactory`, `CursorFactory.canVectorize`, `CursorFactory.makeCursors`, and `CursorFactory.makeVectorCursor`
* updated all `StorageAdapter` implementations to implement `makeCursorHolder`, transitioned direct `CursorFactory` implementations to instead implement `CursorMakerFactory`. `StorageAdapter` being a `CursorMakerFactory` is intended to be a transitional thing, ideally will not be released in favor of moving `CursorMakerFactory` to be fetched directly from `Segment`, however this PR was already large enough so this will be done in a follow-up.
* updated all query engines to use `makeCursorHolder`, granularity based engines to use `CursorGranularizer`.
Two performance enhancements:
1) Direct merging of input frames to output channels, without any
temporary files, if all input frames fit in memory.
2) When doing multi-level merging (now called "external mode"),
improve parallelism by boosting up the number of mergers in the
penultimate level.
To support direct merging, FrameChannelMerger is enhanced such that the
output partition min/max values are used to filter input frames. This
is necessary because all direct mergers read all input frames, but only
rows corresponding to a single output partition.
Some general refactors across Druid.
Switch to DruidExceptions
Add javadocs
Fix a bug in IntArrayColumns
Add a class for LongArrayColumns
Remove wireTransferable since it would never be called
Refactor DictionaryWriter to return the index written as a return value from write.
Refactors the SemanticCreator annotation.
Moves the interface to the semantic package.
Create a SemanticUtils to hold logic for storing semantic maps.
Add FrameMaker interface.
This PR fixes query correctness issues for MSQ window functions when using more than 1 worker (that is, maxNumTasks > 2).
Currently, we were keeping the shuffle spec of the previous stage when we didn't have any partition columns for window stage. This PR changes it to override the shuffle spec of the previous stage to MixShuffleSpec (if we have a window function with empty over clause) so that the window stage gets a single partition to work on.
A test has been added for a query which returned incorrect results prior to this change when using more than 1 workers.
This patch introduces an optional cluster configuration, druid.indexing.formats.stringMultiValueHandlingMode, allowing operators to override the default mode SORTED_SET for string dimensions. The possible values for the config are SORTED_SET, SORTED_ARRAY, or ARRAY (SORTED_SET is the default). Case insensitive values are allowed.
While this cluster property allows users to manage the multi-value handling mode for string dimension types, it's recommended to migrate to using real array types instead of MVDs.
This fixes a long-standing issue where compaction will honor the configured cluster wide property instead of rewriting it as the default SORTED_ARRAY always, even if the data was originally ingested with ARRAY or SORTED_SET.
* MSQ worker: Support in-memory shuffles.
This patch is a follow-up to #16168, adding worker-side support for
in-memory shuffles. Changes include:
1) Worker-side code now respects the same context parameter "maxConcurrentStages"
that was added to the controller in #16168. The parameter remains undocumented
for now, to give us a chance to more fully develop and test this functionality.
1) WorkerImpl is broken up into WorkerImpl, RunWorkOrder, and RunWorkOrderListener
to improve readability.
2) WorkerImpl has a new StageOutputHolder + StageOutputReader concept, which
abstract over memory-based or file-based stage results.
3) RunWorkOrder is updated to create in-memory stage output channels when
instructed to.
4) ControllerResource is updated to add /doneReadingInput/, so the controller
can tell when workers that sort, but do not gather statistics, are done reading
their inputs.
5) WorkerMemoryParameters is updated to consider maxConcurrentStages.
Additionally, WorkerChatHandler is split into WorkerResource, so as to match
ControllerChatHandler and ControllerResource.
* Updates for static checks, test coverage.
* Fixes.
* Remove exception.
* Changes from review.
* Address static check.
* Changes from review.
* Improvements to docs and method names.
* Update comments, add test.
* Additional javadocs.
* Fix throws.
* Fix worker stopping in tests.
* Fix stuck test.
* Round-robin iterator for datasources to kill.
Currently there's a fairness problem in the KillUnusedSegments duty
where the duty consistently selects the same set of datasources as discovered
from the metadata store or dynamic config params. This is a problem especially
when there are multiple unused. In a medium to large cluster, while we can increase
the task slots to increase the likelihood of broader coverage. This patch adds a simple
round-robin iterator to select datasources and has the following properties:
1. Starts with an initial random cursor position in an ordered list of candidates.
2. Consecutive {@code next()} iterations from {@link #getIterator()} are guaranteed to be deterministic
unless the set of candidates change when {@link #updateCandidates(Set)} is called.
3. Guarantees that no duplicate candidates are returned in two consecutive {@code next()} iterations.
* Renames in RoundRobinIteratorTest.
* Address review comments.
1. Clarify javadocs on the ordered list. Also flesh out the details a bit more.
2. Rename the test hooks to make intent clearer and fix typo.
3. Add NotThreadSafe annotation.
4. Remove one potentially noisy log that's in the path of iteration.
* Add null check to input candidates.
* More commentary.
* Addres review feedback: downgrade some new info logs to debug; invert condition.
Remove redundant comments.
Remove rendundant variable tracking.
* CircularList adjustments.
* Updates to CircularList and cleanup RoundRobinInterator.
* One more case and add more tests.
* Make advanceCursor private for now.
* Review comments.
* Coerce COMPLEX to number in numeric aggregators.
PR #15371 eliminated ObjectColumnSelector's built-in implementations of
numeric methods, which had been marked deprecated.
However, some complex types, like SpectatorHistogram, can be successfully coerced
to number. The documentation for spectator histograms encourages taking advantage of
this by aggregating complex columns with doubleSum and longSum. Currently, this
doesn't work properly for IncrementalIndex, where the behavior relied on those
deprecated ObjectColumnSelector methods.
This patch fixes the behavior by making two changes:
1) SimpleXYZAggregatorFactory (XYZ = type; base class for simple numeric aggregators;
all of these extend NullableNumericAggregatorFactory) use getObject for STRING
and COMPLEX. Previously, getObject was only used for STRING.
2) NullableNumericAggregatorFactory (base class for simple numeric aggregators)
has a new protected method "useGetObject". This allows the base class to
correctly check for null (using getObject or isNull).
The patch also adds a test for SpectatorHistogram + doubleSum + IncrementalIndex.
* Fix tests.
* Remove the special ColumnValueSelector.
* Add test.
* HashJoinEngine: Check for interruptions while walking left cursor.
Previously, the engine only checked for interruptions between emitting
joined rows. In scenarios where large numbers of left rows are skipped
completely (such as a highly selective INNER JOIN) this led to the
join cursor being insufficiently responsive to cancellation.
* Coverage.
Changes the WindowFrame internals / representation a bit; introduces dedicated frametypes for rows and groups which corresponds to the implemented processing methods
* more aggressive cancellation of broker parallel merge, more chill blocking queue timeouts
* wire parallel merge into query cancellation system
* oops
* style
* adjust metrics initialization
* fix timeout, fix cleanup to not block
* javadocs to clarify why cancellation future and gizmo are split
* cancelled -> canceled, simplify QueuePusher since it always takes a ResultBatch, non-static terminal marker to make stuff stop complaining about types, specialize tryOffer to be tryOfferTerminal so it wont be misused, add comments to clarify reason for non-blocking offers that might fail
For aggregators like StringFirst/Last, whose intermediate type isn't the same as the final type, using them in GroupBy, TopN or Timeseries subqueries causes a fallback when maxSubqueryBytes is set. This is because we assume that the finalization is not known, due to which the row signature cannot determine whether to use the intermediate or the final type, and it puts it as null. This PR figures out the finalization from the query context and uses the intermediate or the final type appropriately.
changes:
* moves value column serializer initialization, call to `writeValue` method to `GlobalDictionaryEncodedFieldColumnWriter.writeTo` instead of during `GlobalDictionaryEncodedFieldColumnWriter.addValue`. This shift means these numeric value columns are now done in the per field section that happens after serializing the nested column raw data, so only a single compression buffer and temp file will be needed at a time instead of the total number of nested literal fields present in the column. This should be especially helpful for complicated nested structures with thousands of columns as even those 64k compression buffers can add up pretty quickly to a sizeable chunk of direct memory.
changes:
* removes `druid.indexer.task.batchProcessingMode` in favor of always using `CLOSED_SEGMENT_SINKS` which uses `BatchAppenderator`. This was intended to become the default for native batch, but that was missed so `CLOSED_SEGMENTS` was the default (using `AppenderatorImpl`), however MSQ has been exclusively using `BatchAppenderator` with no problems so it seems safe to just roll it out as the only option for batch ingestion everywhere.
* with `batchProcessingMode` gone, there is no use for `AppenderatorImpl` so it has been removed
* implify `Appenderator` construction since there are only separate stream and batch versions now
* simplify tests since `batchProcessingMode` is gone
This PR aims to check if the complex column being queried aligns with the supported types in the aggregator and aggregator factories, and throws a user-friendly error message if they don't.
* Throw exception if DISTINCT used with window functions aggregate call
* Improve error message when unsupported aggregations are used with window functions
changes:
* removed `Firehose` and `FirehoseFactory` and remaining implementations which were mostly no longer used after #16602
* Moved `IngestSegmentFirehose` which was still used internally by Hadoop ingestion to `DatasourceRecordReader.SegmentReader`
* Rename `SQLFirehoseFactoryDatabaseConnector` to `SQLInputSourceDatabaseConnector` and similar renames for sub-classes
* Moved anything remaining in a 'firehose' package somewhere else
* Clean up docs on firehose stuff
Changes:
- Do not hold a reference to `TaskQueue` in `TaskStorageQueryAdapter`
- Use `TaskStorage` instead of `TaskStorageQueryAdapter` in `IndexerMetadataStorageAdapter`
- Rename `TaskStorageQueryAdapter` to `TaskQueryTool`
- Fix newly added task actions `RetrieveUpgradedFromSegmentIds` and `RetrieveUpgradedToSegmentIds`
by removing `isAudited` method.
Description:
Task action audit logging was first deprecated and disabled by default in Druid 0.13, #6368.
As called out in the original discussion #5859, there are several drawbacks to persisting task action audit logs.
- Only usage of the task audit logs is to serve the API `/indexer/v1/task/{taskId}/segments`
which returns the list of segments created by a task.
- The use case is really narrow and no prod clusters really use this information.
- There can be better ways of obtaining this information, such as the metric
`segment/added/bytes` which reports both the segment ID and task ID
when a segment is committed by a task. We could also include committed segment IDs in task reports.
- A task persisting several segments would bloat up the audit logs table putting unnecessary strain
on metadata storage.
Changes:
- Remove `TaskAuditLogConfig`
- Remove method `TaskAction.isAudited()`. No task action is audited anymore.
- Remove `SegmentInsertAction` as it is not used anymore. `SegmentTransactionalInsertAction`
is the new incarnation which has been in use for a while.
- Deprecate `MetadataStorageActionHandler.addLog()` and `getLogs()`. These are not used anymore
but need to be retained for backward compatibility of extensions.
- Do not create `druid_taskLog` metadata table anymore.
* When an ArrayList RAC creates a child RAC, the start and end offsets need to have the offset of parent's start offset
* Defaults the 2nd window bound to CURRENT ROW when only a single bound is specified
* Removes the windowingStrictValidation warning and throws a hard exception when Order By alongside RANGE clause is not provided with UNBOUNDED or CURRENT ROW as both bounds
Better fallback strategy when the broker is unable to materialize the subquery's results as frames for estimating the bytes:
a. We don't touch the subquery sequence till we know that we can materialize the result as frames
Description:
Compaction operations issued by the Coordinator currently run using the native query engine.
As majority of the advancements that we are making in batch ingestion are in MSQ, it is imperative
that we support compaction on MSQ to make Compaction more robust and possibly faster.
For instance, we have seen OOM errors in native compaction that MSQ could have handled by its
auto-calculation of tuning parameters.
This commit enables compaction on MSQ to remove the dependency on native engine.
Main changes:
* `DataSourceCompactionConfig` now has an additional field `engine` that can be one of
`[native, msq]` with `native` being the default.
* if engine is MSQ, `CompactSegments` duty assigns all available compaction task slots to the
launched `CompactionTask` to ensure full capacity is available to MSQ. This is to avoid stalling which
could happen in case a fraction of the tasks were allotted and they eventually fell short of the number
of tasks required by the MSQ engine to run the compaction.
* `ClientCompactionTaskQuery` has a new field `compactionRunner` with just one `engine` field.
* `CompactionTask` now has `CompactionRunner` interface instance with its implementations
`NativeCompactinRunner` and `MSQCompactionRunner` in the `druid-multi-stage-query` extension.
The objectmapper deserializes `ClientCompactionRunnerInfo` in `ClientCompactionTaskQuery` to the
`CompactionRunner` instance that is mapped to the specified type [`native`, `msq`].
* `CompactTask` uses the `CompactionRunner` instance it receives to create the indexing tasks.
* `CompactionTask` to `MSQControllerTask` conversion logic checks whether metrics are present in
the segment schema. If present, the task is created with a native group-by query; if not, the task is
issued with a scan query. The `storeCompactionState` flag is set in the context.
* Each created `MSQControllerTask` is launched in-place and its `TaskStatus` tracked to determine the
final status of the `CompactionTask`. The id of each of these tasks is the same as that of `CompactionTask`
since otherwise, the workers will be unable to determine the controller task's location for communication
(as they haven't been launched via the overlord).
Gian Merlino
Adarsh Sanjeev
Clint Wylie
Benedict Jin
Sree Charan Manamala
Akshat Jain
Laksh Singla
Abhishek Radhakrishnan
Rohan Garg
Zoltan Haindrich
Alberic Liu
Kashif Faraz
Vishesh Garg