During ingestion, incremental segments are created in memory for the different time chunks and persisted to disk when certain thresholds are reached (max number of rows, max memory, incremental persist period etc). In the case where there are a lot of dimension and metrics (1000+) it was observed that the creation/serialization of incremental segment file format for persistence and persisting the file took a while and it was blocking ingestion of new data. This affected the real-time ingestion. This serialization and persistence can be parallelized across the different time chunks. This update aims to do that.
The patch adds a simple configuration parameter to the ingestion tuning configuration to specify number of persistence threads. The default value is 1 if it not specified which makes it the same as it is today.
This PR wires up ValueIndexes and ArrayElementIndexes for nested arrays, ValueIndexes for nested long and double columns, and fixes a handful of bugs I found after adding nested columns to the filter test gauntlet.
introduce checks to ensure that window frame is supported
added check to ensure that no expressions are set as bounds
added logic to detect following/following like cases - described in Window function fails to demarcate if 2 following are used #15739
currently RANGE frames are only supported correctly if both endpoints are unbounded or current row Offset based window range support #15767
added windowingStrictValidation context key to provide a way to override the check
Adds a set of benchmark queries for measuring the planning time with the IN operator. Current results indicate that with the recent optimizations, the IN planning time with 100K expressions in the IN clause is just 3s and with 1M is 46s. For IN clause paired with OR <col>=<val> expr, the numbers are 10s and 155s for 100K and 1M, resp.
Nested columns maintain a null value bitmap for which rows are nulls, however I forgot to wire up a ColumnIndexSupplier to nested columns when filtering the 'raw' data itself, so these were not able to be used. This PR fixes that by adding a supplier that can return NullValueIndex to be used by the NullFilter, which should speed up is null and is not null filters on json columns.
I haven't spent the time to measure the difference yet, but I imagine it should be a significant speed increase.
Note that I only wired this up if druid.generic.useDefaultValueForNull=false (sql compatible mode), the reason being that the SQL planner still uses selector filter, which is unable to properly handle any arrays or complex types (including json, even checking for nulls). The reason for this is so that the behavior is consistent between using the index and using the value matcher, otherwise we get into a situation where using the index has correct behavior but using the value matcher does not, which I was trying to avoid.
* Possibly stabilize intellij-inspections
* remove `integration-tests-ex/cases` from excluded projects from initial build
* enable ErrorProne's `CheckedExceptionNotThrown` to get earlier errors than intellij-inspections
* fix ddsketch pom.xml
* fix spellcheck
* New: Add DDSketch-Druid extension
- Based off of http://www.vldb.org/pvldb/vol12/p2195-masson.pdf and uses
the corresponding https://github.com/DataDog/sketches-java library
- contains tests for post building and using aggregation/post
aggregation.
- New aggregator: `ddSketch`
- New post aggregators: `quantileFromDDSketch` and
`quantilesFromDDSketch`
* Fixing easy CodeQL warnings/errors
* Fixing docs, and dependencies
Also moved aggregator ids to AggregatorUtil and PostAggregatorIds
* Adding more Docs and better null/empty handling for aggregators
* Fixing docs, and pom version
* DDSketch documentation format and wording
A low value of inSubQueryThreshold can cause queries with IN filter to plan as joins more commonly. However, some of these join queries may not get planned as IN filter on data nodes and causes significant perf regression.
### Description
Our Kinesis consumer works by using the [GetRecords API](https://docs.aws.amazon.com/kinesis/latest/APIReference/API_GetRecords.html) in some number of `fetchThreads`, each fetching some number of records (`recordsPerFetch`) and each inserting into a shared buffer that can hold a `recordBufferSize` number of records. The logic is described in our documentation at: https://druid.apache.org/docs/27.0.0/development/extensions-core/kinesis-ingestion/#determine-fetch-settings
There is a problem with the logic that this pr fixes: the memory limits rely on a hard-coded “estimated record size” that is `10 KB` if `deaggregate: false` and `1 MB` if `deaggregate: true`. There have been cases where a supervisor had `deaggregate: true` set even though it wasn’t needed, leading to under-utilization of memory and poor ingestion performance.
Users don’t always know if their records are aggregated or not. Also, even if they could figure it out, it’s better to not have to. So we’d like to eliminate the `deaggregate` parameter, which means we need to do memory management more adaptively based on the actual record sizes.
We take advantage of the fact that GetRecords doesn’t return more than 10MB (https://docs.aws.amazon.com/streams/latest/dev/service-sizes-and-limits.html ):
This pr:
eliminates `recordsPerFetch`, always use the max limit of 10000 records (the default limit if not set)
eliminate `deaggregate`, always have it true
cap `fetchThreads` to ensure that if each fetch returns the max (`10MB`) then we don't exceed our budget (`100MB` or `5% of heap`). In practice this means `fetchThreads` will never be more than `10`. Tasks usually don't have that many processors available to them anyway, so in practice I don't think this will change the number of threads for too many deployments
add `recordBufferSizeBytes` as a bytes-based limit rather than records-based limit for the shared queue. We do know the byte size of kinesis records by at this point. Default should be `100MB` or `10% of heap`, whichever is smaller.
add `maxBytesPerPoll` as a bytes-based limit for how much data we poll from shared buffer at a time. Default is `1000000` bytes.
deprecate `recordBufferSize`, use `recordBufferSizeBytes` instead. Warning is logged if `recordBufferSize` is specified
deprecate `maxRecordsPerPoll`, use `maxBytesPerPoll` instead. Warning is logged if maxRecordsPerPoll` is specified
Fixed issue that when the record buffer is full, the fetchRecords logic throws away the rest of the GetRecords result after `recordBufferOfferTimeout` and starts a new shard iterator. This seems excessively churny. Instead, wait an unbounded amount of time for queue to stop being full. If the queue remains full, we’ll end up right back waiting for it after the restarted fetch.
There was also a call to `newQ::offer` without check in `filterBufferAndResetBackgroundFetch`, which seemed like it could cause data loss. Now checking return value here, and failing if false.
### Release Note
Kinesis ingestion memory tuning config has been greatly simplified, and a more adaptive approach is now taken for the configuration. Here is a summary of the changes made:
eliminates `recordsPerFetch`, always use the max limit of 10000 records (the default limit if not set)
eliminate `deaggregate`, always have it true
cap `fetchThreads` to ensure that if each fetch returns the max (`10MB`) then we don't exceed our budget (`100MB` or `5% of heap`). In practice this means `fetchThreads` will never be more than `10`. Tasks usually don't have that many processors available to them anyway, so in practice I don't think this will change the number of threads for too many deployments
add `recordBufferSizeBytes` as a bytes-based limit rather than records-based limit for the shared queue. We do know the byte size of kinesis records by at this point. Default should be `100MB` or `10% of heap`, whichever is smaller.
add `maxBytesPerPoll` as a bytes-based limit for how much data we poll from shared buffer at a time. Default is `1000000` bytes.
deprecate `recordBufferSize`, use `recordBufferSizeBytes` instead. Warning is logged if `recordBufferSize` is specified
deprecate `maxRecordsPerPoll`, use `maxBytesPerPoll` instead. Warning is logged if maxRecordsPerPoll` is specified
* IncrementalIndex#add is no longer thread-safe.
Following #14866, there is no longer a reason for IncrementalIndex#add
to be thread-safe.
It turns out it already was not using its selectors in a thread-safe way,
as exposed by #15615 making `testMultithreadAddFactsUsingExpressionAndJavaScript`
in `IncrementalIndexIngestionTest` flaky. Note that this problem isn't
new: Strings have been stored in the dimension selectors for some time,
but we didn't have a test that checked for that case; we only have
this test that checks for concurrent adds involving numeric selectors.
At any rate, this patch changes OnheapIncrementalIndex to no longer try
to offer a thread-safe "add" method. It also improves performance a bit
by adding a row ID supplier to the selectors it uses to read InputRows,
meaning that it can get the benefit of caching values inside the selectors.
This patch also:
1) Adds synchronization to HyperUniquesAggregator and CardinalityAggregator,
which the similar datasketches versions already have. This is done to
help them adhere to the contract of Aggregator: concurrent calls to
"aggregate" and "get" must be thread-safe.
2) Updates OnHeapIncrementalIndexBenchmark to use JMH and moves it to the
druid-benchmarks module.
* Spelling.
* Changes from static analysis.
* Fix javadoc.
* Clear "lineSplittable" for JSON when using KafkaInputFormat.
JsonInputFormat has a "withLineSplittable" method that can be used to
control whether JSON is read line-by-line, or as a whole. The intent
is that in streaming ingestion, "lineSplittable" is false (although it
can be overridden by "assumeNewlineDelimited"), and in batch ingestion,
lineSplittable is true.
When a "json" format is wrapped by a "kafka" format, this isn't set
properly. This patch updates KafkaInputFormat to set this on an
underlying "json" format.
The tests for KafkaInputFormat were overriding the "lineSplittable"
parameter explicitly, which wasn't really fair, because that made them
unrealistic to what happens in production. Now they omit the parameter
and get the production behavior.
* Add test.
* Fix test coverage.
* Faster parsing: reduce String usage, list-based input rows.
Three changes:
1) Reworked FastLineIterator to optionally avoid generating Strings
entirely, and reduce copying somewhat. Benefits the line-oriented
JSON, CSV, delimited (TSV), and regex formats.
2) In the delimited (TSV) format, when the delimiter is a single byte,
split on UTF-8 bytes directly.
3) In CSV and delimited (TSV) formats, use list-based input rows when
the column list is provided upfront by the user.
* Fix style.
* Fix inspections.
* Restore validation.
* Remove fastutil-extra.
* Exception type.
* Fixes for error messages.
* Fixes for null handling.
This PR fixes the summary iterator to add aggregators in the correct position. The summary iterator is used when dims are not present, therefore the new change is identical to the old one, but seems more correct while reading.
* support groups windowing mode; which is a close relative of ranges (but not in the standard)
* all windows with range expressions will be executed wit it groups
* it will be 100% correct in case for both bounds its true that: isCurrentRow() || isUnBounded()
* this covers OVER ( ORDER BY COL )
* for other cases it will have some chances of getting correct results...
* Cache value selectors in RowBasedColumnSelectorFactory.
There was already caching for dimension selectors. This patch adds caching
for value (object and number) selectors. It's helpful when the same field is
read multiple times during processing of a single row (for example, by being
an input to both MIN and MAX aggregations).
* Fix typing.
* Fix logic.
* Add SpectatorHistogram extension
* Clarify documentation
Cleanup comments
* Use ColumnValueSelector directly
so that we support being queried as a Number using longSum or doubleSum aggregators as well as a histogram.
When queried as a Number, we're returning the count of entries in the histogram.
* Apply suggestions from code review
Co-authored-by: Victoria Lim <vtlim@users.noreply.github.com>
* Fix references
* Fix spelling
* Update docs/development/extensions-contrib/spectator-histogram.md
Co-authored-by: Victoria Lim <vtlim@users.noreply.github.com>
---------
Co-authored-by: Victoria Lim <vtlim@users.noreply.github.com>
* Add ImmutableLookupMap for static lookups.
This patch adds a new ImmutableLookupMap, which comes with an
ImmutableLookupExtractor. It uses a fastutil open hashmap plus two
lists to store its data in such a way that forward and reverse
lookups can both be done quickly. I also observed footprint to be
somewhat smaller than Java HashMap + MapLookupExtractor for a 1 million
row lookup.
The main advantage, though, is that reverse lookups can be done much
more quickly than MapLookupExtractor (which iterates the entire map
for each call to unapplyAll). This speeds up the recently added
ReverseLookupRule (#15626) during SQL planning with very large lookups.
* Use in one more test.
* Fix benchmark.
* Object2ObjectOpenHashMap
* Fixes, and LookupExtractor interface update to have asMap.
* Remove commented-out code.
* Fix style.
* Fix import order.
* Add fastutil.
* Avoid storing Map entries.
* Reverse, pull up lookups in the SQL planner.
Adds two new rules:
1) ReverseLookupRule, which eliminates calls to LOOKUP by doing
reverse lookups.
2) AggregatePullUpLookupRule, which pulls up calls to LOOKUP above
GROUP BY, when the lookup is injective.
Adds configs `sqlReverseLookup` and `sqlPullUpLookup` to control whether
these rules fire. Both are enabled by default.
To minimize the chance of performance problems due to many keys mapping to
the same value, ReverseLookupRule refrains from reversing a lookup if there
are more keys than `inSubQueryThreshold`. The rationale for using this setting
is that reversal works by generating an IN, and the `inSubQueryThreshold`
describes the largest IN the user wants the planner to create.
* Add additional line.
* Style.
* Remove commented-out lines.
* Fix tests.
* Add test.
* Fix doc link.
* Fix docs.
* Add one more test.
* Fix tests.
* Logic, test updates.
* - Make FilterDecomposeConcatRule more flexible.
- Make CalciteRulesManager apply reduction rules til fixpoint.
* Additional tests, simplify code.
* Faster k-way merging using tournament trees, 8-byte key strides.
Two speedups for FrameChannelMerger (which does k-way merging in MSQ):
1) Replace the priority queue with a tournament tree, which does fewer
comparisons.
2) Compare keys using 8-byte strides, rather than 1 byte at a time.
* Adjust comments.
* Fix style.
* Adjust benchmark and test.
* Add eight-list test (power of two).
This PR fixes a bug with the long string pair serde where null and empty strings are treated equivalently, and the return value is always null. When 'useDefaultValueForNull' was set to true by default, this wasn't a commonly seen issue, because nulls were equivalent to empty strings. However, since the default has changed to false, this can create incorrect results when the long string pairs are serded, where the empty strings are incorrectly converted to nulls.
changes:
* ColumnIndexSelector now extends ColumnSelector. The only real implementation of ColumnIndexSelector, ColumnSelectorColumnIndexSelector, already has a ColumnSelector, so this isn't very disruptive
* removed getColumnNames from ColumnSelector since it was not used
* VirtualColumns and VirtualColumn getIndexSupplier method now needs argument of ColumnIndexSelector instead of ColumnSelector, which allows expression virtual columns to correctly recognize other virtual columns, fixing an issue which would incorrectly handle other virtual columns as non-existent columns instead
* fixed a bug with sql planner incorrectly not using expression filter for equality filters on columns with extractionFn and no virtual column registry
* overhaul DruidPredicateFactory to better handle 3VL
fixes some bugs caused by some limitations of the original design of how DruidPredicateFactory interacts with 3-value logic. The primary impacted area was with how filters on values transformed with expressions or extractionFn which turn non-null values into nulls, which were not possible to be modelled with the 'isNullInputUnknown' method
changes:
* adds DruidObjectPredicate to specialize string, array, and object based predicates instead of using guava Predicate
* DruidPredicateFactory now uses DruidObjectPredicate
* introduces DruidPredicateMatch enum, which all predicates returned from DruidPredicateFactory now use instead of booleans to indicate match. This means DruidLongPredicate, DruidFloatPredicate, DruidDoublePredicate, and the newly added DruidObjectPredicate apply methods all now return DruidPredicateMatch. This allows matchers and indexes
* isNullInputUnknown has been removed from DruidPredicateFactory
* rename, fix test
* adjust
* style
* npe
* more test
* fix default value mode to not match new test
* Reverse lookup fixes and enhancements.
1) Add a "mayIncludeUnknown" parameter to DimFilter#optimize. This is important
because otherwise the reverse-lookup optimization is done improperly when
the "in" filter appears under a "not", and the lookup extractionFn may return
null for some possible values of the filtered column. The "includeUnknown" test
cases in InDimFilterTest illustrate the difference in behavior.
2) Enhance InDimFilter#optimizeLookup to handle "mayIncludeUnknown", and to be able
to do a reverse lookup in a wider variety of cases.
3) Make "unapply" protected in LookupExtractor, and move callers to "unapplyAll".
The main reason is that MapLookupExtractor, a common implementation, lacks a
reverse mapping and therefore does a scan of the map for each call to "unapply".
For performance sake these calls need to be batched.
* Remove optimize call from BloomDimFilter.
* Follow the law.
* Fix tests.
* Fix imports.
* Switch function.
* Fix tests.
* More tests.
* New handling for COALESCE, SEARCH, and filter optimization.
COALESCE is converted by Calcite's parser to CASE, which is largely
counterproductive for us, because it ends up duplicating expressions.
In the current code we end up un-doing it in our CaseOperatorConversion.
This patch has a different approach:
1) Add CaseToCoalesceRule to convert CASE back to COALESCE earlier, before
the Volcano planner runs, using CaseToCoalesceRule.
2) Add FilterDecomposeCoalesceRule to decompose calls like
"f(COALESCE(x, y))" into "(x IS NOT NULL AND f(x)) OR (x IS NULL AND f(y))".
This helps use indexes when available on x and y.
3) Add CoalesceLookupRule to push COALESCE into the third arg of LOOKUP.
4) Add a native "coalesce" function so we can convert 3+ arg COALESCE.
The advantage of this approach is that by un-doing the CASE to COALESCE
conversion earlier, we have flexibility to do more stuff with
COALESCE (like decomposition and pushing into LOOKUP).
SEARCH is an operator used internally by Calcite to represent matching
an argument against some set of ranges. This patch improves our handling
of SEARCH in two ways:
1) Expand NOT points (point "holes" in the range set) from SEARCH as
`!(a || b)` rather than `!a && !b`, which makes it possible to convert
them to a "not" of "in" filter later.
2) Generate those nice conversions for NOT points even if the SEARCH
is not composed of 100% NOT points. Without this change, a SEARCH
for "x NOT IN ('a', 'b') AND x < 'm'" would get converted like
"x < 'a' OR (x > 'a' AND x < 'b') OR (x > 'b' AND x < 'm')".
One of the steps we take when generating Druid queries from Calcite
plans is to optimize native filters. This patch improves this step:
1) Extract common ANDed predicates in ConvertSelectorsToIns, so we can
convert "(a && x = 'b') || (a && x = 'c')" into "a && x IN ('b', 'c')".
2) Speed up CombineAndSimplifyBounds and ConvertSelectorsToIns on
ORs with lots of children by adjusting the logic to avoid calling
"indexOf" and "remove" on an ArrayList.
3) Refactor ConvertSelectorsToIns to reduce duplicated code between the
handling for "selector" and "equals" filters.
* Not so final.
* Fixes.
* Fix test.
* Fix test.
* Fix ColumnSelectorColumnIndexSelector#getColumnCapabilities.
It was using virtualColumns.getColumnCapabilities, which only returns
capabilities for virtual columns, not regular columns. The effect of this
is that expression filters (and in some cases, arrayContainsElement filters)
would build value matchers rather than use indexes.
I think this has been like this since #12315, which added the
getColumnCapabilities method to BitmapIndexSelector, and included the same
implementation as exists in the code today.
This error is easy to make due to the design of virtualColumns.getColumnCapabilities,
so to help avoid it in the future, this patch renames the method to
getColumnCapabilitiesWithoutFallback to emphasize that it does not return
capabilities for regular columns.
* Make getColumnCapabilitiesWithoutFallback package-private.
* Fix expression filter bitmap usage.
The PR: #13947 introduced a function evalDimension() in the interface RowFunction.
There was no default implementation added for this interface which causes all the implementations and custom transforms to fail and require to implement their own version of evalDimension method. This PR adds a default implementation in the interface which allows the evalDimension to return value as a Singleton array of eval result.
Fixes#15072
Before this modification , the third parameter (timezone) require to be a Literal, it will throw a error when this parameter is column Identifier.
Changes
- Add `log` implementation for `AuditManager` alongwith `SQLAuditManager`
- `LoggingAuditManager` simply logs the audit event. Thus, it returns empty for
all `fetchAuditHistory` calls.
- Add new config `druid.audit.manager.type` which can take values `log`, `sql` (default)
- Add new config `druid.audit.manager.logLevel` which can take values `DEBUG`, `INFO`, `WARN`.
This gets activated only if `type` is `log`.
- Remove usage of `ConfigSerde` from `AuditManager` as audit is not just limited to configs
- Add `AuditSerdeHelper` for a single implementation of serialization/deserialization of
audit payload and other utility methods.
* Allow for kafka emitter producer secrets to be masked in logs instead of being visible
This change will allow for kafka producer config values that should be secrets to not show up in the logs.
This will enhance the security of the people who use the kafka emitter to use this if they want to.
This is opt in and will not affect prior configs for this emitter
* fix checkstyle issue
* change property name
I was looking into a query which was performing a bit poorly because the case_searched was touching more than 1 columns (if there is only 1 column there is a cache based evaluator).
While I was doing that I've noticed that there are a few simple things which could help a bit:
use a static TRUE/FALSE instead of creating a new object every time
create the ExprEval early for ConstantExpr -s (except the one for BigInteger which seem to have some odd contract)
return early from type autodetection
these changes mostly reduce the amount of garbage the query creates during case_searched evaluation; although ExpressionSelectorBenchmark shows some improvements ~15% - but my manual trials on the taxi dataset with 60M rows showed more improvements - probably due to the fact that these changes mostly only reduce gc pressure.
Pramod Immaneni
Soumyava
Clint Wylie
Zoltan Haindrich
Vishesh Garg
Abhishek Agarwal
Abhishek Radhakrishnan
Gian Merlino
Hiroshi Fukada
Karan Kumar
Pranav
zachjsh
Benedict Jin
Laksh Singla
AmatyaAvadhanula
Sam Rash
Kashif Faraz
Ben Sykes
Parth Agrawal
AlbericByte
Tom