fix: Rewrite Spark materialization engine to use mapInPandas #3936
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: tokoko <togurg14@freeuni.edu.ge>
HaoXuAI
approved these changes
Feb 8, 2024
|
Unfortunately we aren't testing data types extensively, so that's really hard to say. There will definitely be some conversion changes between pyspark<->pandas translation, because current implementation goes through python dicts and discards numeric types smaller than int for example. Having said that, it's unlikely to affect the final feast data types as those aren't part of feast type system anyway. |
|
@tokoko can you please resolve the conflicts and I'll add the |
|
@HaoXuAI All fixed, thanks. |
|
Curious if you are able to test the performance somehow? |
|
Not really, this one seemed too obvious to do that. I plan to later try out conditionally using |
tqtensor
pushed a commit
to tqtensor/feast
that referenced
this pull request
Mar 11, 2024
…ev#3936) rewrite spark materilization engine to use mapInPandas Signed-off-by: tokoko <togurg14@freeuni.edu.ge>
|
+1 it also seems like this should be using real Spark sink connectors rather than ad-hoc procedural side effects. |
|
@rehevkor5 true, but that's hard to do if you want to support arbitrary online store implementations. Maybe we should keep this as the default way to materialize and somehow introduce a way to override the logic for specific online stores. Not sure where that sort of code belongs to, though... online stores or materilization engines? |
franciscojavierarceo
pushed a commit
that referenced
this pull request
Apr 16, 2024
# [0.36.0](v0.35.0...v0.36.0) (2024-04-16) ### Bug Fixes * Add __eq__, __hash__ to SparkSource for correct comparison ([#4028](#4028)) ([e703b40](e703b40)) * Add conn.commit() to Postgresonline_write_batch.online_write_batch ([#3904](#3904)) ([7d75fc5](7d75fc5)) * Add missing __init__.py to embedded_go ([#4051](#4051)) ([6bb4c73](6bb4c73)) * Add missing init files in infra utils ([#4067](#4067)) ([54910a1](54910a1)) * Added registryPath parameter documentation in WebUI reference ([#3983](#3983)) ([5e0af8f](5e0af8f)), closes [#3974](#3974) [#3974](#3974) * Adding missing init files in materialization modules ([#4052](#4052)) ([df05253](df05253)) * Allow trancated timestamps when converting ([#3861](#3861)) ([bdd7dfb](bdd7dfb)) * Azure blob storage support in Java feature server ([#2319](#2319)) ([#4014](#4014)) ([b9aabbd](b9aabbd)) * Bugfix for grabbing historical data from Snowflake with array type features. ([#3964](#3964)) ([1cc94f2](1cc94f2)) * Bytewax materialization engine fails when loading feature_store.yaml ([#3912](#3912)) ([987f0fd](987f0fd)) * CI unittest warnings ([#4006](#4006)) ([0441b8b](0441b8b)) * Correct the returning class proto type of StreamFeatureView to StreamFeatureViewProto instead of FeatureViewProto. ([#3843](#3843)) ([86d6221](86d6221)) * Create index only if not exists during MySQL online store update ([#3905](#3905)) ([2f99a61](2f99a61)) * Disable minio tests in workflows on master and nightly ([#4072](#4072)) ([c06dda8](c06dda8)) * Disable the Feast Usage feature by default. ([#4090](#4090)) ([b5a7013](b5a7013)) * Dump repo_config by alias ([#4063](#4063)) ([e4bef67](e4bef67)) * Extend SQL registry config with a sqlalchemy_config_kwargs key ([#3997](#3997)) ([21931d5](21931d5)) * Feature Server image startup in OpenShift clusters ([#4096](#4096)) ([9efb243](9efb243)) * Fix copy method for StreamFeatureView ([#3951](#3951)) ([cf06704](cf06704)) * Fix for materializing entityless feature views in Snowflake ([#3961](#3961)) ([1e64c77](1e64c77)) * Fix type mapping spark ([#4071](#4071)) ([3afa78e](3afa78e)) * Fix typo as the cli does not support shortcut-f option. ([#3954](#3954)) ([dd79dbb](dd79dbb)) * Get container host addresses from testcontainers ([#3946](#3946)) ([2cf1a0f](2cf1a0f)) * Handle ComplexFeastType to None comparison ([#3876](#3876)) ([fa8492d](fa8492d)) * Hashlib md5 errors in FIPS for python 3.9+ ([#4019](#4019)) ([6d9156b](6d9156b)) * Making the query_timeout variable as optional int because upstream is considered to be optional ([#4092](#4092)) ([fd5b620](fd5b620)) * Move gRPC dependencies to an extra ([#3900](#3900)) ([f93c5fd](f93c5fd)) * Prevent spamming pull busybox from dockerhub ([#3923](#3923)) ([7153cad](7153cad)) * Quickstart notebook example ([#3976](#3976)) ([b023aa5](b023aa5)) * Raise error when not able read of file source spark source ([#4005](#4005)) ([34cabfb](34cabfb)) * remove not use input parameter in spark source ([#3980](#3980)) ([7c90882](7c90882)) * Remove parentheses in pull_latest_from_table_or_query ([#4026](#4026)) ([dc4671e](dc4671e)) * Remove proto-plus imports ([#4044](#4044)) ([ad8f572](ad8f572)) * Remove unnecessary dependency on mysqlclient ([#3925](#3925)) ([f494f02](f494f02)) * Restore label check for all actions using pull_request_target ([#3978](#3978)) ([591ba4e](591ba4e)) * Revert mypy config ([#3952](#3952)) ([6b8e96c](6b8e96c)) * Rewrite Spark materialization engine to use mapInPandas ([#3936](#3936)) ([dbb59ba](dbb59ba)) * Run feature server w/o gunicorn on windows ([#4024](#4024)) ([584e9b1](584e9b1)) * SqlRegistry _apply_object update statement ([#4042](#4042)) ([ef62def](ef62def)) * Substrait ODFVs for online ([#4064](#4064)) ([26391b0](26391b0)) * Swap security label check on the PR title validation job to explicit permissions instead ([#3987](#3987)) ([f604af9](f604af9)) * Transformation server doesn't generate files from proto ([#3902](#3902)) ([d3a2a45](d3a2a45)) * Trino as an OfflineStore Access Denied when BasicAuthenticaion ([#3898](#3898)) ([49d2988](49d2988)) * Trying to import pyspark lazily to avoid the dependency on the library ([#4091](#4091)) ([a05cdbc](a05cdbc)) * Typo Correction in Feast UI Readme ([#3939](#3939)) ([c16e5af](c16e5af)) * Update actions/setup-python from v3 to v4 ([#4003](#4003)) ([ee4c4f1](ee4c4f1)) * Update typeguard version to >=4.0.0 ([#3837](#3837)) ([dd96150](dd96150)) * Upgrade sqlalchemy from 1.x to 2.x regarding PVE-2022-51668. ([#4065](#4065)) ([ec4c15c](ec4c15c)) * Use CopyFrom() instead of __deepycopy__() for creating a copy of protobuf object. ([#3999](#3999)) ([5561b30](5561b30)) * Using version args to install the correct feast version ([#3953](#3953)) ([b83a702](b83a702)) * Verify the existence of Registry tables in snowflake before calling CREATE sql command. Allow read-only user to call feast apply. ([#3851](#3851)) ([9a3590e](9a3590e)) ### Features * Add duckdb offline store ([#3981](#3981)) ([161547b](161547b)) * Add Entity df in format of a Spark Dataframe instead of just pd.DataFrame or string for SparkOfflineStore ([#3988](#3988)) ([43b2c28](43b2c28)) * Add gRPC Registry Server ([#3924](#3924)) ([373e624](373e624)) * Add local tests for s3 registry using minio ([#4029](#4029)) ([d82d1ec](d82d1ec)) * Add python bytes to array type conversion support proto ([#3874](#3874)) ([8688acd](8688acd)) * Add python client for remote registry server ([#3941](#3941)) ([42a7b81](42a7b81)) * Add Substrait-based ODFV transformation ([#3969](#3969)) ([9e58bd4](9e58bd4)) * Add support for arrays in snowflake ([#3769](#3769)) ([8d6bec8](8d6bec8)) * Added delete_table to redis online store ([#3857](#3857)) ([03dae13](03dae13)) * Adding support for Native Python feature transformations for ODFVs ([#4045](#4045)) ([73bc853](73bc853)) * Bumping requirements ([#4079](#4079)) ([1943056](1943056)) * Decouple transformation types from ODFVs ([#3949](#3949)) ([0a9fae8](0a9fae8)) * Dropping Python 3.8 from local integration tests and integration tests ([#3994](#3994)) ([817995c](817995c)) * Dropping python 3.8 requirements files from the project. ([#4021](#4021)) ([f09c612](f09c612)) * Dropping the support for python 3.8 version from feast ([#4010](#4010)) ([a0f7472](a0f7472)) * Dropping unit tests for Python 3.8 ([#3989](#3989)) ([60f24f9](60f24f9)) * Enable Arrow-based columnar data transfers ([#3996](#3996)) ([d8d7567](d8d7567)) * Enable Vector database and retrieve_online_documents API ([#4061](#4061)) ([ec19036](ec19036)) * Kubernetes materialization engine written based on bytewax ([#4087](#4087)) ([7617bdb](7617bdb)) * Lint with ruff ([#4043](#4043)) ([7f1557b](7f1557b)) * Make arrow primary interchange for offline ODFV execution ([#4083](#4083)) ([9ed0a09](9ed0a09)) * Pandas v2 compatibility ([#3957](#3957)) ([64459ad](64459ad)) * Pull duckdb from contribs, add to CI ([#4059](#4059)) ([318a2b8](318a2b8)) * Refactor ODFV schema inference ([#4076](#4076)) ([c50a9ff](c50a9ff)) * Refactor registry caching logic into a separate class ([#3943](#3943)) ([924f944](924f944)) * Rename OnDemandTransformations to Transformations ([#4038](#4038)) ([9b98eaf](9b98eaf)) * Revert updating dependencies so that feast can be run on 3.11. ([#3968](#3968)) ([d3c68fb](d3c68fb)), closes [#3958](#3958) * Rewrite ibis point-in-time-join w/o feast abstractions ([#4023](#4023)) ([3980e0c](3980e0c)) * Support s3gov schema by snowflake offline store during materialization ([#3891](#3891)) ([ea8ad17](ea8ad17)) * Update odfv test ([#4054](#4054)) ([afd52b8](afd52b8)) * Update pyproject.toml to use Python 3.9 as default ([#4011](#4011)) ([277b891](277b891)) * Update the Pydantic from v1 to v2 ([#3948](#3948)) ([ec11a7c](ec11a7c)) * Updating dependencies so that feast can be run on 3.11. ([#3958](#3958)) ([59639db](59639db)) * Updating protos to separate transformation ([#4018](#4018)) ([c58ef74](c58ef74)) ### Reverts * Reverting bumping requirements ([#4081](#4081)) ([1ba65b4](1ba65b4)), closes [#4079](#4079) * Verify the existence of Registry tables in snowflake… ([#3907](#3907)) ([c0d358a](c0d358a)), closes [#3851](#3851)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What this PR does / why we need it:
Spark materialization engine currently converts data on partitions to a list of dicts before trying to write it to online store. This is terribly inefficient and all but guarantees OOM errors on executors for sufficiently large datasets. This PR changes the implementation to use
mapInPandasand sidesteps the conversion to a list of dicts entirely.P.S. There's unfortunately no equivalent
applyInPandasmethod in pyspark, so the implementation simulates it by usingmapInPandasand returning dummy results back to the driver.P.P.S. Starting from 3.3.0 pyspark also has
mapInArrow, which would simplify the operation even more, but would mean that feast would no longer support pyspark versions older than 3.3.0. I decided not to use it for now.