* Lower default maxRowsInMemory for realtime ingestion.
The thinking here is that for best ingestion throughput, we want
intermediate persists to be as big as possible without using up all
available memory. So, we rely mainly on maxBytesInMemory. The default
maxRowsInMemory (1 million) is really just a safety: in case we have
a large number of very small rows, we don't want to get overwhelmed
by per-row overheads.
However, maximum ingestion throughput isn't necessarily the primary
goal for realtime ingestion. Query performance is also important. And
because query performance is not as good on the in-memory dataset, it's
helpful to keep it from growing too large. 150k seems like a reasonable
balance here. It means that for a typical 5 million row segment, we
won't trigger more than 33 persists due to this limit, which is a
reasonable number of persists.
* Update tests.
* Update server/src/main/java/org/apache/druid/segment/indexing/RealtimeTuningConfig.java
Co-authored-by: Kashif Faraz <kashif.faraz@gmail.com>
* Fix test.
* Fix link.
---------
Co-authored-by: Kashif Faraz <kashif.faraz@gmail.com>
* Removing intermediateSuperSorterStorageMaxLocalBytes, maxInputBytesPerWorker, composedIntermediateSuperSorterStorageEnabled, clusterStatisticsMergeMode from docs
* Adding documentation in the context class.
* Various changes and fixes to UNNEST.
Native changes:
1) UnnestDataSource: Replace "column" and "outputName" with "virtualColumn".
This enables pushing expressions into the datasource. This in turn
allows us to do the next thing...
2) UnnestStorageAdapter: Logically apply query-level filters and virtual
columns after the unnest operation. (Physically, filters are pulled up,
when possible.) This is beneficial because it allows filters and
virtual columns to reference the unnested column, and because it is
consistent with how the join datasource works.
3) Various documentation updates, including declaring "unnest" as an
experimental feature for now.
SQL changes:
1) Rename DruidUnnestRel (& Rule) to DruidUnnestRel (& Rule). The rel
is simplified: it only handles the UNNEST part of a correlated join.
Constant UNNESTs are handled with regular inline rels.
2) Rework DruidCorrelateUnnestRule to focus on pulling Projects from
the left side up above the Correlate. New test testUnnestTwice verifies
that this works even when two UNNESTs are stacked on the same table.
3) Include ProjectCorrelateTransposeRule from Calcite to encourage
pushing mappings down below the left-hand side of the Correlate.
4) Add a new CorrelateFilterLTransposeRule and CorrelateFilterRTransposeRule
to handle pulling Filters up above the Correlate. New tests
testUnnestWithFiltersOutside and testUnnestTwiceWithFilters verify
this behavior.
5) Require a context feature flag for SQL UNNEST, since it's undocumented.
As part of this, also cleaned up how we handle feature flags in SQL.
They're now hooked into EngineFeatures, which is useful because not
all engines support all features.
* Sort-merge join and hash shuffles for MSQ.
The main changes are in the processing, multi-stage-query, and sql modules.
processing module:
1) Rename SortColumn to KeyColumn, replace boolean descending with KeyOrder.
This makes it nicer to model hash keys, which use KeyOrder.NONE.
2) Add nullability checkers to the FieldReader interface, and an
"isPartiallyNullKey" method to FrameComparisonWidget. The join
processor uses this to detect null keys.
3) Add WritableFrameChannel.isClosed and OutputChannel.isReadableChannelReady
so callers can tell which OutputChannels are ready for reading and which
aren't.
4) Specialize FrameProcessors.makeCursor to return FrameCursor, a random-access
implementation. The join processor uses this to rewind when it needs to
replay a set of rows with a particular key.
5) Add MemoryAllocatorFactory, which is embedded inside FrameWriterFactory
instead of a particular MemoryAllocator. This allows FrameWriterFactory
to be shared in more scenarios.
multi-stage-query module:
1) ShuffleSpec: Add hash-based shuffles. New enum ShuffleKind helps callers
figure out what kind of shuffle is happening. The change from SortColumn
to KeyColumn allows ClusterBy to be used for both hash-based and sort-based
shuffling.
2) WorkerImpl: Add ability to handle hash-based shuffles. Refactor the logic
to be more readable by moving the work-order-running code to the inner
class RunWorkOrder, and the shuffle-pipeline-building code to the inner
class ShufflePipelineBuilder.
3) Add SortMergeJoinFrameProcessor and factory.
4) WorkerMemoryParameters: Adjust logic to reserve space for output frames
for hash partitioning. (We need one frame per partition.)
sql module:
1) Add sqlJoinAlgorithm context parameter; can be "broadcast" or
"sortMerge". With native, it must always be "broadcast", or it's a
validation error. MSQ supports both. Default is "broadcast" in
both engines.
2) Validate that MSQs do not use broadcast join with RIGHT or FULL join,
as results are not correct for broadcast join with those types. Allow
this in native for two reasons: legacy (the docs caution against it,
but it's always been allowed), and the fact that it actually *does*
generate correct results in native when the join is processed on the
Broker. It is much less likely that MSQ will plan in such a way that
generates correct results.
3) Remove subquery penalty in DruidJoinQueryRel when using sort-merge
join, because subqueries are always required, so there's no reason
to penalize them.
4) Move previously-disabled join reordering and manipulation rules to
FANCY_JOIN_RULES, and enable them when using sort-merge join. Helps
get to better plans where projections and filters are pushed down.
* Work around compiler problem.
* Updates from static analysis.
* Fix @param tag.
* Fix declared exception.
* Fix spelling.
* Minor adjustments.
* wip
* Merge fixups
* fixes
* Fix CalciteSelectQueryMSQTest
* Empty keys are sortable.
* Address comments from code review. Rename mux -> mix.
* Restore inspection config.
* Restore original doc.
* Reorder imports.
* Adjustments
* Fix.
* Fix imports.
* Adjustments from review.
* Update header.
* Adjust docs.
You can now do the following operations with TupleSketches in Post Aggregation Step
Get the Sketch Output as Base64 String
Provide a constant Tuple Sketch in post-aggregation step that can be used in Set Operations
Get the Estimated Value(Sum) of Summary/Metrics Objects associated with Tuple Sketch
The FiniteFirehoseFactory and InputRowParser classes were deprecated in 0.17.0 (#8823) in favor of InputSource & InputFormat. This PR removes the FiniteFirehoseFactory and all its implementations along with classes solely used by them like Fetcher (Used by PrefetchableTextFilesFirehoseFactory). Refactors classes including tests using FiniteFirehoseFactory to use InputSource instead.
Removing InputRowParser may not be as trivial as many classes that aren't deprecated depends on it (with no alternatives), like EventReceiverFirehoseFactory. Hence FirehoseFactory, EventReceiverFirehoseFactory, and Firehose are marked deprecated.
*When running REPLACE queries, the segments which contain no data are dropped (marked as unused). This PR aims to generate tombstones in place of segments which contain no data to mark their deletion, as is the behavior with the native ingestion.
This will cause InsertCannotReplaceExistingSegmentFault to be removed since it was generated if the interval to be marked unused didn't fully overlap one of the existing segments to replace.
* Fix NPE in KinesisSupervisor#setupRecordSupplier.
PR #13539 refactored record supplier creation and introduced a bug:
this method would throw NPE when recordsPerFetch was not provided
by the user. recordsPerFetch isn't needed in this context at all,
since the supervisor-side supplier doesn't fetch records. So this
patch sets it to zero.
* Remove unused imports.
If the intermediate handoff period is less than the task duration and there is no new data in the input topic, task will continuously checkpoint the same offsets again and again. This PR fixes that bug by resetting the checkpoint time even when the task receives the same end offset request again.
* merge druid-core, extendedset, and druid-hll into druid-processing to simplify everything
* fix poms and license stuff
* mockito is evil
* allow reset of JvmUtils RuntimeInfo if tests used static injection to override
* Use an HllSketchHolder object to enable optimized merge
HllSketchAggregatorFactory.combine had been implemented using a
pure pair-wise, "make a union -> add 2 things to union -> get sketch"
algorithm. This algorithm does 2 things that was CPU
1) The Union object always builds an HLL_8 sketch regardless of the
target type. This means that when the target type is not HLL_8, we
spent CPU cycles converting to HLL_8 and back over and over again
2) By throwing away the Union object and converting back to the
HllSketch only to build another Union object, we do lots and lots
of copy+conversions of the HllSketch
This change introduces an HllSketchHolder object which can hold onto
a Union object and delay conversion back into an HllSketch until
it is actually needed. This follows the same pattern as the
SketchHolder object for theta sketches.
changes:
* modified druid schema column type compution to special case COMPLEX<json> handling to choose COMPLEX<json> if any column in any segment is COMPLEX<json>
* NestedFieldVirtualColumn can now work correctly on any type of column, returning either a column selector if a root path, or nil selector if not
* fixed a random bug with NilVectorSelector when using a vector size larger than the default and druid.generic.useDefaultValueForNull=false would have the nulls vector set to all false instead of true
* fixed an overly aggressive check in ExprEval.ofType when handling complex types which would try to treat any string as base64 without gracefully falling back if it was not in fact base64 encoded, along with special handling for complex<json>
* added ExpressionVectorSelectors.castValueSelectorToObject and ExpressionVectorSelectors.castObjectSelectorToNumeric as convience methods to cast vector selectors using cast expressions without the trouble of constructing an expression. the polymorphic nature of the non-vectorized engine (and significantly larger overhead of non-vectorized expression processing) made adding similar methods for non-vectorized selectors less attractive and so have not been added at this time
* fix inconsistency between nested column indexer and serializer in handling values (coerce non primitive and non arrays of primitives using asString)
* ExprEval best effort mode now handles byte[] as string
* added test for ExprEval.bestEffortOf, and add missing conversion cases that tests uncovered
* more tests more better
* Fallback virtual column
This virtual columns enables falling back to another column if
the original column doesn't exist. This is useful when doing
column migrations and you have some old data with column X,
new data with column Y and you want to use Y if it exists, X
otherwise so that you can run a consistent query against all of
the data.
With fault tolerance enabled in MSQ, not all the work orders might be populated if the worker is restarted. In case it gets the request for cleaning up the stage which is not present in the worker's map, it can throw an NPE. Added a check to ensure that the stage is present in the map before cleaning it up, or else logging it as a warning.
* SQL test framework extensions
* Capture planner artifacts: logical plan, etc.
* Planner test builder validates the logical plan
* Validation for the SQL resut schema (we already have
validation for the Druid row signature)
* Better Guice integration: properties, reuse Guice modules
* Avoid need for hand-coded expr, macro tables
* Retire some of the test-specific query component creation
* Fix query log hook race condition
Co-authored-by: Paul Rogers <progers@apache.org>
* discover nested columns when using nested column indexer for schemaless
* move useNestedColumnIndexerForSchemaDiscovery from AppendableIndexSpec to DimensionsSpec
Much improved table functions
* Revises properties, definitions in the catalog
* Adds a "table function" abstraction to model such functions
* Specific functions for HTTP, inline, local and S3.
* Extended SQL types in the catalog
* Restructure external table definitions to use table functions
* EXTEND syntax for Druid's extern table function
* Support for array-valued table function parameters
* Support for array-valued SQL query parameters
* Much new documentation
* Kinesis: More robust default fetch settings.
1) Default recordsPerFetch and recordBufferSize based on available memory
rather than using hardcoded numbers. For this, we need an estimate
of record size. Use 10 KB for regular records and 1 MB for aggregated
records. With 1 GB heaps, 2 processors per task, and nonaggregated
records, recordBufferSize comes out to the same as the old
default (10000), and recordsPerFetch comes out slightly lower (1250
instead of 4000).
2) Default maxRecordsPerPoll based on whether records are aggregated
or not (100 if not aggregated, 1 if aggregated). Prior default was 100.
3) Default fetchThreads based on processors divided by task count on
Indexers, rather than overall processor count.
4) Additionally clean up the serialized JSON a bit by adding various
JsonInclude annotations.
* Updates for tests.
* Additional important verify.
* Quote and escape table, key and column names.
* fix typo.
* More select statements.
* Derby lookup tests create quoted identifiers so it's compatible.
* Use Stringutils.replace() utility.
* quote the filter string.
* Squish doubly quote usage into a single function.
* Add parameterized test with reserved identifiers.
* few changes.
* Addition of NaiveSortMaker and Default implementation
Add the NaiveSortMaker which makes a sorter
object and a default implementation of the
interface.
This also allows us to plan multiple different window
definitions on the same query.
* Validate response headers and fix exception logging
A class of QueryException were throwing away their
causes making it really hard to determine what's
going wrong when something goes wrong in the SQL
planner specifically. Fix that and adjust tests
to do more validation of response headers as well.
We allow 404s and 307s to be returned even without
authorization validated, but others get converted to 403
This PR expands `StringDimensionIndexer` to handle conversion of `byte[]` to base64 encoded strings, rather than the current behavior of calling java `toString`.
This issue was uncovered by a regression of sorts introduced by #13519, which updated the protobuf extension to directly convert stuff to java types, resulting in `bytes` typed values being converted as `byte[]` instead of a base64 string which the previous JSON based conversion created. While outputting `byte[]` is more consistent with other input formats, and preferable when the bytes can be consumed directly (such as complex types serde), when fed to a `StringDimensionIndexer`, it resulted in an ugly java `toString` because `processRowValsToUnsortedEncodedKeyComponent` is fed the output of `row.getRaw(..)`. Converting `byte[]` to a base64 string within `StringDimensionIndexer` is consistent with the behavior of calling `row.getDimension(..)` which does do this coercion (and why many tests on binary types appeared to be doing the expected thing).
I added some protobuf `bytes` tests, but they don't really hit the new `StringDimensionIndexer` behavior because they operate on the `InputRow` directly, and call `getDimension` to validate stuff. The parser based version still uses the old conversion mechanisms, so when not using a flattener incorrectly calls `toString` on the `ByteString`. I have encoded this behavior in the test for now, if we either update the parser to use the new flattener or just .. remove parsers we can remove this test stuff.
Gian Merlino
Adarsh Sanjeev
Karan Kumar
Tejaswini Bandlamudi
Laksh Singla
Clint Wylie
Rohan Garg
Nicholas Lippis
Anshu Makkar
Paul Rogers
Abhishek Agarwal
Abhishek Radhakrishnan
imply-cheddar
somu-imply
Maytas Monsereenusorn
Kashif Faraz