* fix MAX_INTERMEDIATE_SIZE for DoubleMeanHolder
* byte[] type handling in deserialize and finalizeComputation for DoubleMeanAggregatorFactory
* DoubleMeanAggregatorFactory tests: Max Intermediate Size, Deserialize, finalizeComputation
* moved byte[] check to first position
Co-authored-by: Stanislav <S.Poryadnyi@abcconsulting.ru>
* fix nullhandling exceptions related to test ordering
Tests might get executed in different order depending on the maven
version and the test environment. This may lead to "NullHandling module
not initialized" errors for some tests where we do not initialize
null-handling explicitly.
* use InitializedNullHandlingTest
* SQL support for joins on subqueries.
Changes to SQL module:
- DruidJoinRule: Allow joins on subqueries (left/right are no longer
required to be scans or mappings).
- DruidJoinRel: Add cost estimation code for joins on subqueries.
- DruidSemiJoinRule, DruidSemiJoinRel: Removed, since DruidJoinRule can
handle this case now.
- DruidRel: Remove Nullable annotation from toDruidQuery, because
it is no longer needed (it was used by DruidSemiJoinRel).
- Update Rules constants to reflect new rules available in our current
version of Calcite. Some of these are useful for optimizing joins on
subqueries.
- Rework cost estimation to be in terms of cost per row, and place all
relevant constants in CostEstimates.
Other changes:
- RowBasedColumnSelectorFactory: Don't set hasMultipleValues. The lack
of isComplete is enough to let callers know that columns might have
multiple values, and explicitly setting it to true causes
ExpressionSelectors to think it definitely has multiple values, and
treat the inputs as arrays. This behavior interfered with some of the
new tests that involved queries on lookups.
- QueryContexts: Add maxSubqueryRows parameter, and use it in druid-sql
tests.
* Fixes for tests.
* Adjustments.
* Broker: Add ability to inline subqueries.
The main changes:
- ClientQuerySegmentWalker: Add ability to inline queries.
- Query: Add "getSubQueryId" and "withSubQueryId" methods.
- QueryMetrics: Add "subQueryId" dimension.
- ServerConfig: Add new "maxSubqueryRows" parameter, which is used by
ClientQuerySegmentWalker to limit how many rows can be inlined per
query.
- IndexedTableJoinMatcher: Allow creating keys on top of unknown types,
by assuming they are strings. This is useful because not all types are
known for fields in query results.
- InlineDataSource: Store RowSignature rather than component parts. Add
more zealous "equals" and "hashCode" methods to ease testing.
- Moved QuerySegmentWalker test code from CalciteTests and
SpecificSegmentsQueryWalker in druid-sql to QueryStackTests in
druid-server. Use this to spin up a new ClientQuerySegmentWalkerTest.
* Adjustments from CI.
* Fix integration test.
* Move RowSignature from druid-sql to druid-processing and make use of it.
1) Moved (most of) RowSignature from sql to processing. Left behind the SQL-specific
stuff in a RowSignatures utility class. It also picked up some new convenience
methods along the way.
2) There were a lot of places in the code where Map<String, ValueType> was used to
associate columns with type info. These are now all replaced with RowSignature.
3) QueryToolChest's resultArrayFields method is replaced with resultArraySignature,
and it now provides type info.
* Fix up extensions.
* Various fixes
* Link up row-based datasources to serving layer.
- Add SegmentWrangler interface that allows linking of DataSources to Segments.
- Add LocalQuerySegmentWalker that uses SegmentWranglers to compute queries on
data that is available locally.
- Modify ClientQuerySegmentWalker to use LocalQuerySegmentWalker when the base
datasource is concrete and not a table.
- Add SegmentWranglerModule to the Broker so it has them available and can
properly instantiate . LocalQuerySegmentWalkers.
- Set InlineDataSource and LookupDataSource to concrete, since they can be
directly queried now.
* Fix tests.
* Ability to directly query row-based datasources.
Includes:
- Foundational classes RowBasedSegment, RowBasedStorageAdapter,
RowBasedCursor provide a queryable interface on top of a
RowBasedColumnSelectorFactory.
- Add LookupSegment: A RowBasedSegment that is built on lookup data.
- Improve capability reporting in RowBasedColumnSelectorFactory.
* Fix import.
* Remove unthrown IOException.
* Harmonization and bug-fixing for selector and filter behavior on unknown types.
- Migrate ValueMatcherColumnSelectorStrategy to newer ColumnProcessorFactory
system, and set defaultType COMPLEX so unknown types can be dynamically matched.
- Remove ValueGetters in favor of ColumnComparisonFilter doing its own thing.
- Switch various methods to use convertObjectToX when casting to numbers, rather
than ad-hoc and inconsistent logic.
- Fix bug in RowBasedExpressionColumnValueSelector: isBindingArray should return
true even for 0- or 1- element arrays.
- Adjust various javadocs.
* Add throwParseExceptions option to Rows.objectToNumber, switch back to that.
* Update tests.
* Adjust moment sketch tests.
* Add OnHeapMemorySegmentWriteOutMediumFactory
Add a factory for OnHeapMemorySegmentWriteOutMedium to support direct writing via Spark.
* Register OnHeapMemorySegmentWriteOutMediumFactory.
Register OnHeapMemorySegmentWriteOutMediumFactory with SegmentWriteOutMediumFactory.
* Remove unnecessary throws
The base `makeSegmentWriteOutMedium` throws an IOException, but the particular implementation of OnHeapMemorySegmentWriteOutMediumFactory does not throw a checked exception.
* Update SegmentWriteOutMedium docs to include onHeapMemory
Update the SegmentWriteOutMedium section of the indexing docs to include a description of the new OnHeapSegmentMediumWriteOut option.
* BufferArrayGrouper: Fix potential overflow in requiredBufferCapacity.
If cardinality was high, the computation could overflow an int. There
were tests for this, but the tests were wrong.
* Nicer.
* Add SQL GROUPING SETS support.
Built on top of the subtotalsSpec feature in the groupBy query. This also involves
two changes to subtotalsSpec:
- Alter behavior so limitSpec is applied after subtotalsSpec, rather than applied to
each grouping set. This is more in line with SQL standard behavior. I think it is okay
to make this change, since the old behavior was not documented, so users should
hopefully not be depending on it.
- Fix a bug where virtual columns were included in the subtotal queries, but they
should not have been.
Also fixes two bugs in query equality checking:
- BaseQuery: Use getDuration() instead of "duration" in equals and hashCode, since the
latter is lazily initialized and might be null in one query but not the other.
- GroupByQuery: Include subtotalsSpec in equals and hashCode.
* Fix bugs.
* Fix tests.
* PR updates.
* Grouping class hygiene.
* add Expr.stringify which produces parseable expression strings, parser support for null values in arrays, and parser support for empty numeric arrays
* oops, macros are expressions too
* style
* spotbugs
* qualified type arrays
* review stuffs
* simplify grammar
* more permissive array parsing
* reuse expr joiner
* fix it
* Run IntelliJ inspections on Travis
Running IntelliJ inspections currently takes about 90 minutes, but they
can be run in about 30 minutes on Travis.
* Restore assert statements
* Fix timestamp extract fn to match postgres
Update the timestamp extract function so that it matches the PostgreSQL docs.
Examples from the PostgreSQL docs were added as tests for DECADE, CENTURY
and MILLENIUM extraction.
There were bugs in CENTURY and MILLENIUM that were spotted because of intelliJ
inspections - 'Integer division in floating point context'
* Update CalciteQueryTest
* remove useless round
* mark integer division as an error
When the time extraction Top N algorithm is looking for aggregators, it makes
2 calls to hashCode on the key. Use Map#computeIfAbsent instead so that the
hashCode is calculated only once
* Codestyle - use java style array declaration
Replaced C-style array declarations with java style declarations and marked
the intelliJ inspection as an error
* cleanup test code
* Forbid easily misused HashSet and HashMap constructors
* Add two LinkedHashMap constructors to forbidden-apis and create utility method as replacement for them
* Fix visibility of constant in CollectionUtils.java
* Make an exception for an instance of LinkedHashMap#<init>(int) because proper sizing is used
* revert changes to sql module tests that should be in separate PR
* Finish reverting changes to sql module tests that were flagged in checkstyle during CI
* Add netty dependency resulting from SupressForbidden
* Add HashVectorGrouper based on MemoryOpenHashTable.
Additional supporting changes:
1) Modifies VectorGrouper interface to use Memory instead of ByteBuffers.
2) Modifies BufferArrayGrouper to match the new VectorGrouper interface.
3) Removes "implements VectorGrouper" from BufferHashGrouper.
* Fix comment.
* Fix another comment.
* Remove unused stuff.
* Include hoisted bounds checks.
* Checks against too-large keySpaces.
* Add MemoryOpenHashTable, a table similar to ByteBufferHashTable.
With some key differences to improve speed and design simplicity:
1) Uses Memory rather than ByteBuffer for its backing storage.
2) Uses faster hashing and comparison routines (see HashTableUtils).
3) Capacity is always a power of two, allowing simpler design and more
efficient implementation of findBucket.
4) Does not implement growability; instead, leaves that to its callers.
The idea is this removes the need for subclasses, while still giving
callers flexibility in how to handle table-full scenarios.
* Fix LGTM warnings.
* Adjust dependencies.
* Remove easymock from druid-benchmarks.
* Adjustments from review.
* Fix datasketches unit tests.
* Fix checkstyle.
* Speed up joins on indexed tables with string keys
When joining on index tables with string keys, caching the computation
of row id to row numbers improves performance on the
JoinAndLookupBenchmark.joinIndexTableStringKey* benchmarks by about 10%
if the column cache is enabled an by about 100% if the column cache is
disabled.
* Faster cache impl and handle unknown cardinality
* Remove unused dependency
* Hoist cardinality check outside of hot loop
* Fix dummy DimensionSelector for tests
* Guicify druid sql module
Break up the SQLModule in to smaller modules and provide a binding that
modules can use to register schemas with druid sql.
* fix some tests
* address code review
* tests compile
* Working tests
* Add all the tests
* fix up licenses and dependencies
* add calcite dependency to druid-benchmarks
* tests pass
* rename the schemas
* SQL join support for lookups.
1) Add LookupSchema to SQL, so lookups show up in the catalog.
2) Add join-related rels and rules to SQL, allowing joins to be planned into
native Druid queries.
* Add two missing LookupSchema calls in tests.
* Fix tests.
* Fix typo.
* Add LookupJoinableFactory.
Enables joins where the right-hand side is a lookup. Includes an
integration test.
Also, includes changes to LookupExtractorFactoryContainerProvider:
1) Add "getAllLookupNames", which will be needed to eventually connect
lookups to Druid's SQL catalog.
2) Convert "get" from nullable to Optional return.
3) Swap out most usages of LookupReferencesManager in favor of the
simpler LookupExtractorFactoryContainerProvider interface.
* Fixes for tests.
* Fix another test.
* Java 11 message fix.
* Fixups.
* Fixup benchmark class.
* Add getRightColumns to JoinConditionAnalysis
This change other implementations of JoinableFactory to ask the analysis
for the right key columns instead of having to calculate it themselves.
* Address some review comments
* more code review stuff
Add microbenchmark for joins. Enabling the column cache improves
performance by ~70% for the benchmarks for joins with string keys.
Adjusting LookupJoinMatcher.matchCondition() to have fewer branches,
improves performance by ~10% for the benchmarks for joins with lookups.
* intelliJ inspections cleanup
- remove redundant escapes
- performance warnings
- access static member via instance reference
- static method declared final
- inner class may be static
Most of these changes are aesthetic, however, they will allow inspections to
be enabled as part of CI checks going forward
The valuable changes in this delta are:
- using StringBuilder instead of string addition in a loop
indexing-hadoop/.../Utils.java
processing/.../ByteBufferMinMaxOffsetHeap.java
- Use class variables instead of static variables for parameterized test
processing/src/.../ScanQueryLimitRowIteratorTest.java
* Add intelliJ inspection warnings as errors to druid profile
* one more static inner class
* Make JoinableFactory an extension point
This change makes it so that extensions can register a JoinableFactory that
should be used for a DataSource.
Extensions can provide the factories via DruidBinders#joinableFactoryBinder
Known DataSources - like InlineDataSource are provided in the
JoinableFactoryModule. This module installs a FactoryWarehouse that is
used to decide which factory should be used to generate the Joinable for
the provided DataSource.
The ExtensionPoint is marked as Beta since it is not yet clear if this
needs to remain available to other extensions or if the best way to
register a factory is by using the datasource class.
* Add module test
* remove useless bindings in test
* remove ExtensionPoint annotation
* Make LifecycleLock not final to help with testing
* Add JoinableFactory interface and use it in the query stack.
Also includes InlineJoinableFactory, which enables joining against
inline datasources. This is the first patch where a basic join query
actually works. It includes integration tests.
* Fix test issues.
* Adjustments from code review.
Builds on #9235, using the datasource analysis functionality to replace various ad-hoc
approaches. The most interesting changes are in ClientQuerySegmentWalker (brokers),
ServerManager (historicals), and SinkQuerySegmentWalker (indexing tasks).
Other changes related to improving how we analyze queries:
1) Changes TimelineServerView to return an Optional timeline, which I thought made
the analysis changes cleaner to implement.
2) Added QueryToolChest#canPerformSubquery, which is now used by query entry points to
determine whether it is safe to pass a subquery dataSource to the query toolchest.
Fixes an issue introduced in #5471 where subqueries under non-groupBy-typed queries
were silently ignored, since neither the query entry point nor the toolchest did
anything special with them.
3) Removes the QueryPlus.withQuerySegmentSpec method, which was mostly being used in
error-prone ways (ignoring any potential subqueries, and not verifying that the
underlying data source is actually a table). Replaces with a new function,
Queries.withSpecificSegments, that includes sanity checks.
* Add join-related DataSource types, and analysis functionality.
Builds on #9111 and implements the datasource analysis mentioned in #8728. Still can't
handle join datasources, but we're a step closer.
Join-related DataSource types:
1) Add "join", "lookup", and "inline" datasources.
2) Add "getChildren" and "withChildren" methods to DataSource, which will be used
in the future for query rewriting (e.g. inlining of subqueries).
DataSource analysis functionality:
1) Add DataSourceAnalysis class, which breaks down datasources into three components:
outer queries, a base datasource (left-most of the highest level left-leaning join
tree), and other joined-in leaf datasources (the right-hand branches of the
left-leaning join tree).
2) Add "isConcrete", "isGlobal", and "isCacheable" methods to DataSource in order to
support analysis.
Other notes:
1) Renamed DataSource#getNames to DataSource#getTableNames, which I think is clearer.
Also, made it a Set, so implementations don't need to worry about duplicates.
2) The addition of "isCacheable" should work around #8713, since UnionDataSource now
returns false for cacheability.
* Remove javadoc comment.
* Updates reflecting code review.
* Add comments.
* Add more comments.
* Optimize JoinCondition matching
The LookupJoinMatcher needs to check if a condition is always true or false
multiple times. This can be pre-computed to speed up the match checking
This change reduces the time it takes to perform a for joining on a long key
from ~ 36 ms/op to 23 ms/ op
* Rename variables
* fix typo
* null handling for numeric first/last aggregators, refactor to not extend nullable numeric agg since they are complex typed aggs
* initially null or not based on config
* review stuff, make string first/last consistent with null handling of numeric columns, more tests
* docs
* handle nil selectors, revert to primitive first/last types so groupby v1 works...
* Speed up String first/last aggregators when folding isn't needed.
Examines the value column, and disables fold checking via a needsFoldCheck
flag if that column can't possibly contain SerializableLongStringPairs. This
is helpful because it avoids calling getObject on the value selector when
unnecessary; say, because the time selector didn't yield an earlier or later
value.
* PR comments.
* Move fastLooseChop to StringUtils.
* Add HashJoinSegment, a virtual segment for joins.
An initial step towards #8728. This patch adds enough functionality to implement a joining
cursor on top of a normal datasource. It does not include enough to actually do a query. For
that, future patches will need to wire this low-level functionality into the query language.
* Fixups.
* Fix missing format argument.
* Various tests and minor improvements.
* Changes.
* Remove or add tests for unused stuff.
* Fix up package locations.
* Fix double-checked locking in predicate suppliers in BoundDimFilter
* Fix double-checked locking in predicate suppliers in BoundDimFilter
* 1. Use Suppliers.memoize() to initialize and publish singleton.
2. Fix coding style.
* Fix coding style
* Fix double-checked locking bug for predicate suppliers in InDimFilter
* Add FileUtils.createTempDir() and enforce its usage.
The purpose of this is to improve error messages. Previously, the error
message on a nonexistent or unwritable temp directory would be
"Failed to create directory within 10,000 attempts".
* Further updates.
* Another update.
* Remove commons-io from benchmark.
* Fix tests.
* Refactor parallel indexing perfect rollup partitioning
Refactoring to make it easier to later add range partitioning for
perfect rollup parallel indexing. This is accomplished by adding several
new base classes (e.g., PerfectRollupWorkerTask) and new classes for
encapsulating logic that needs to be changed for different partitioning
strategies (e.g., IndexTaskInputRowIteratorBuilder).
The code is functionally equivalent to before except for the following
small behavior changes:
1) PartialSegmentMergeTask: Previously, this task had a priority of
DEFAULT_TASK_PRIORITY. It now has a priority of
DEFAULT_BATCH_INDEX_TASK_PRIORITY (via the new PerfectRollupWorkerTask
base class), since it is a batch index task.
2) ParallelIndexPhaseRunner: A decorator was added to
subTaskSpecIterator to ensure the subtasks are generated with unique
ids. Previously, only tests (i.e., MultiPhaseParallelIndexingTest)
would have this decorator, but this behavior is desired for non-test
code as well.
* Fix forbidden apis and pmd warnings
* Fix analyze dependencies warnings
* Fix IndexTask json and add IT diags
* Fix parallel index supervisor<->worker serde
* Fix TeamCity inspection errors/warnings
* Fix TeamCity inspection errors/warnings again
* Integrate changes with those from #8823
* Address review comments
* Address more review comments
* Fix forbidden apis
* Address more review comments
* Tidy up lifecycle, query, and ingestion logging.
The goal of this patch is to improve the clarity and usefulness of
Druid's logging for cluster operators. For more information, see
https://twitter.com/cowtowncoder/status/1195469299814555648.
Concretely, this patch does the following:
- Changes a lot of INFO logs to DEBUG, and DEBUG to TRACE, with the
goal of reducing redundancy and improving clarity by avoiding
showing rarely-useful log messages. This includes most "starting"
and "stopping" messages, and most messages related to individual
columns.
- Adds new log4j2 templates that show operators how to enabled DEBUG
logging for certain important packages.
- Eliminate stack traces for query errors, unless log level is DEBUG
or more. This is useful because query errors often indicate user
error rather than system error, but dumping stack trace often gave
operators the impression that there was a system failure.
- Adds task id to Appenderator, AppenderatorDriver thread names. In
the default log4j2 configuration, this will put them in log lines
as well. It's very useful if a user is using the Indexer, where
multiple tasks run in the same JVM.
- More consistent terminology when it comes to "sequences" (sets of
segments that are handed-off together by Kafka ingestion) and
"offsets" (cursors in partitions). These terms had been confused in
some log messages due to the fact that Kinesis calls offsets
"sequence numbers".
- Replaces some ugly toString calls with either the JSONification or
something more operator-accessible (like a URL or segment identifier,
instead of JSON object representing the same).
* Adjustments.
* Adjust integration test.
* transformSpec + array expressions
changes:
* added array expression support to transformSpec
* removed ParseSpec.verify since its only use afaict was preventing transform expr that did not replace their input from functioning
* hijacked index task test to test changes
* remove docs about being unsupported
* re-arrange test assert
* unused imports
* imports
* fix tests
* preserve types
* suppress warning, fixes, add test
* formatting
* cleanup
* better list to array type conversion and tests
* fix oops
* use peekable iterator for numeric column selector null checking instead of bitmap.get for those sweet sweet nanoseconds
* remove unused method
* slight optimization i think
* remove clone from wrappers since we do not use and is confusing
* fixes and tests
* int instead of Integer
* fix it
* fixes, more tests
* fix
* SQL: EARLIEST, LATEST aggregators.
I chose these names instead of FIRST, LAST because those are already
reserved functions in Calcite that mean something different. I think
these are also better names anyway.
* Finalify.
* SQL updates.
* Adjust aggregator calls.
* Validations, test updates.
* Review docs.
There is a class of bugs due to the fact that BaseObjectColumnValueSelector
has both "getObject" and "isNull" methods, but in most selector implementations
and most call sites, it is clear that the intent of "isNull" is only to apply
to the primitive getters, not the object getter. This makes sense, because the
purpose of isNull is to enable detection of nulls in otherwise-primitive columns.
Imagine a string column with a numeric selector built on top of it. You would
want it to return isNull = true, so numeric aggregators don't treat it as
all zeroes.
Sometimes this design leads people to accidentally guard non-primitive get
methods with "selector.isNull" checks, which is improper.
This patch has three goals:
1) Fix null-handling bugs that already exist in this class.
2) Make interface and doc changes that reduce the probability of future bugs.
3) Fix other, unrelated bugs I noticed in the stringFirst and stringLast
aggregators while fixing null-handling bugs. I thought about splitting this
into its own patch, but it ended up being tough to split from the
null-handling fixes.
For (1) the fixes are,
- Fix StringFirst and StringLastAggregatorFactory to stop guarding getObject
calls on isNull, by no longer extending NullableAggregatorFactory. Now uses
-1 as a sigil value for null, to differentiate nulls and empty strings.
- Fix ExpressionFilter to stop guarding getObject calls on isNull. Also, use
eval.asBoolean() to avoid calling getLong on the selector after already
calling getObject.
- Fix ObjectBloomFilterAggregator to stop guarding DimensionSelector calls
on isNull. Also, refactored slightly to avoid the overhead of calling
getObject followed by another getter (see BloomFilterAggregatorFactory for
part of this).
For (2) the main changes are,
- Remove the "isNull" method from BaseObjectColumnValueSelector.
- Clarify "isNull" doc on BaseNullableColumnValueSelector.
- Rename NullableAggregatorFactory -> NullbleNumericAggregatorFactory to emphasize
that it only works on aggregators that take numbers as input.
- Similar naming changes to the Aggregator, BufferAggregator, and AggregateCombiner.
- Similar naming changes to helper methods for groupBy, ValueMatchers, etc.
For (3) the other fixes for StringFirst and StringLastAggregatorFactory are,
- Fixed buffer overrun in the buffer aggregators when some characters in the string
code into more than one byte (the old code used "substring" to apply a byte limit,
which is bad). I did this by introducing a new StringUtils.toUtf8WithLimit method.
- Fixed weird IncrementalIndex logic that led to reading nulls for the timestamp.
- Adjusted weird StringFirst/Last logic that worked around the weird IncrementalIndex
behavior.
- Refactored to share code between the four aggregators.
- Improved test coverage.
- Made the base stringFirst, stringLast aggregators adaptive, and streamlined the
xFold versions into aliases. The adaptiveness is similar to how other aggregators
like hyperUnique work.
* sketch of broker parallel merges done in small batches on fork join pool
* fix non-terminating sequences, auto compute parallelism
* adjust benches
* adjust benchmarks
* now hella more faster, fixed dumb
* fix
* remove comments
* log.info for debug
* javadoc
* safer block for sequence to yielder conversion
* refactor LifecycleForkJoinPool into LifecycleForkJoinPoolProvider which wraps a ForkJoinPool
* smooth yield rate adjustment, more logs to help tune
* cleanup, less logs
* error handling, bug fixes, on by default, more parallel, more tests
* remove unused var
* comments
* timeboundary mergeFn
* simplify, more javadoc
* formatting
* pushdown config
* use nanos consistently, move logs back to debug level, bit more javadoc
* static terminal result batch
* javadoc for nullability of createMergeFn
* cleanup
* oops
* fix race, add docs
* spelling, remove todo, add unhandled exception log
* cleanup, revert unintended change
* another unintended change
* review stuff
* add ParallelMergeCombiningSequenceBenchmark, fixes
* hyper-threading is the enemy
* fix initial start delay, lol
* parallelism computer now balances partition sizes to partition counts using sqrt of sequence count instead of sequence count by 2
* fix those important style issues with the benchmarks code
* lazy sequence creation for benchmarks
* more benchmark comments
* stable sequence generation time
* update defaults to use 100ms target time, 4096 batch size, 16384 initial yield, also update user docs
* add jmh thread based benchmarks, cleanup some stuff
* oops
* style
* add spread to jmh thread benchmark start range, more comments to benchmarks parameters and purpose
* retool benchmark to allow modeling more typical heterogenous heavy workloads
* spelling
* fix
* refactor benchmarks
* formatting
* docs
* add maxThreadStartDelay parameter to threaded benchmark
* why does catch need to be on its own line but else doesnt
Jihoon Son
Abhishek Radhakrishnan
Himanshu
Stanislav Poryadnyi
Xavier Léauté
Clint Wylie
Gian Merlino
Jonathan Wei
Samarth Jain
Julian Jaffe
Chi Cao Minh
Suneet Saldanha
Maytas Monsereenusorn
Lucas Capistrant
Q
Fangyuan Deng
jon-wei
Atul Mohan