[SPARK-42539][SQL][HIVE] Eliminate separate classloader when using 'builtin' Hive version for metadata client#40144
Closed
xkrogen wants to merge 5 commits into
Conversation
…e general spark user classloader when 'builtin' is used instead of reconstructing a new URL classloader
Contributor
Author
|
cc @sunchao @AngersZhuuuu since you've worked on somewhat-related changes in #34690, #32887, etc. |
Member
|
Also, cc @mridulm , @cloud-fan , @rednaxelafx, @zsxwing, @kiszk , @maropu . |
Contributor
|
It makes sense to use the builtin classloader when using builtin Hive. To clarify: we still have the class loading issue if people specifies a certain hive version (not builtin), right? |
Contributor
Author
|
Great question @cloud-fan , and actually no, we don't. For all of the other values of |
xkrogen
changed the title
cloud-fan
approved these changes
Feb 27, 2023
Contributor
|
lgtm if all tests pass |
sunchao
approved these changes
Feb 27, 2023
sunchao
pushed a commit
that referenced
this pull request
Feb 27, 2023
…uiltin' Hive version for metadata client ### What changes were proposed in this pull request? When using the 'builtin' Hive version for the Hive metadata client, do not create a separate classloader, and rather continue to use the overall user/application classloader (regardless of Java version). This standardizes the behavior for all Java versions with that of Java 9+. See SPARK-42539 for more details on why this approach was chosen. ### Why are the changes needed? Please see a much more detailed description in SPARK-42539. The tl;dr is that user-provided JARs (such as `hive-exec-2.3.8.jar`) take precedence over Spark/system JARs when constructing the classloader used by `IsolatedClientLoader` on Java 8 in 'builtin' mode, which can cause unexpected behavior and/or breakages. This violates the expectation that, unless user-first classloader mode is used, Spark JARs should be prioritized over user JARs. It also seems that this separate classloader was unnecessary from the start, since the intent of 'builtin' mode is to use the JARs already existing on the regular classloader (as alluded to [here](#24057 (comment))). The isolated clientloader was originally added in #5876 in 2015. This bit in the PR description is the only mention of the behavior for "builtin": > attempt to discover the jars that were used to load Spark SQL and use those. This option is only valid when using the execution version of Hive. I can't follow the logic here; the user classloader clearly has all of the necessary Hive JARs, since that's where we're getting the JAR URLs from, so we could just use that directly instead of grabbing the URLs. When this was initially added, it only used the JARs from the user classloader, not any of its parents, which I suspect was the motivating factor (to try to avoid more Spark classes being duplicated inside of the isolated classloader, I guess). But that was changed a month later anyway in #6435 / #6459, so I think this may have basically been deadcode from the start. It has also caused at least one issue over the years, e.g. SPARK-21428, which disables the new-classloader behavior in the case of running inside of a CLI session. ### Does this PR introduce _any_ user-facing change? No, except to protect Spark itself from potentially being broken by bad user JARs. ### How was this patch tested? This includes a new unit test in `HiveUtilsSuite` which demonstrates the issue and shows that this approach resolves it. It has also been tested on a live cluster running Java 8 and Hive communication functionality continues to work as expected. Closes #40144 from xkrogen/xkrogen/SPARK-42539/hive-isolatedclientloader-builtin-user-jar-conflict-fix/java9strategy. Authored-by: Erik Krogen <xkrogen@apache.org> Signed-off-by: Chao Sun <sunchao@apple.com>
Member
|
Merged to master/branch-3.4. Thanks @xkrogen ! |
xkrogen
deleted the
xkrogen/SPARK-42539/hive-isolatedclientloader-builtin-user-jar-conflict-fix/java9strategy
branch
Contributor
Author
|
Thanks @sunchao and @cloud-fan ! |
Member
|
Seems like the tests didn't pass .. I am reverting this as it causes a lot of test failures. e.g.) |
Member
|
Thank you for recovering |
Contributor
|
Shall we revert it from 3.4 as well? |
Member
|
Yes, it was reverted here, 26009d4. |
Member
|
Hmm interesting. Somehow the tests were shown all passing for me when I merged this. Sorry for the trouble. |
Contributor
Author
|
I will take a look at the test failures, thanks @HyukjinKwon for addressing the revert! |
Member
|
If you folks don't mind, shall we consider this for Apache Spark 3.5 only? |
Member
|
+1. I agree with @dongjoon-hyun . |
Contributor
Author
|
I agree as well. Posted a new PR at #40224. |
sunchao
pushed a commit
that referenced
this pull request
Mar 1, 2023
…uiltin' Hive version for metadata client ### What changes were proposed in this pull request? When using the 'builtin' Hive version for the Hive metadata client, do not create a separate classloader, and rather continue to use the overall user/application classloader (regardless of Java version). This standardizes the behavior for all Java versions with that of Java 9+. See SPARK-42539 for more details on why this approach was chosen. Please note that this is a re-submit of #40144. That one introduced test failures, and potentially a real issue, because the PR works by setting `isolationOn = false` for `builtin` mode. In addition to adjusting the classloader, `HiveClientImpl` relies on `isolationOn` to determine if it should use an isolated copy of `SessionState`, so the PR inadvertently switched to using a shared `SessionState` object. I think we do want to continue to have the isolated session state even in `builtin` mode, so this adds a new flag `sessionStateIsolationOn` which controls whether the session state should be isolated, _separately_ from the `isolationOn` flag which controls whether the classloader should be isolated. Default behavior is for `sessionStateIsolationOn` to be set equal to `isolationOn`, but for `builtin` mode, we override it to enable session state isolated even though classloader isolation is turned off. ### Why are the changes needed? Please see a much more detailed description in SPARK-42539. The tl;dr is that user-provided JARs (such as `hive-exec-2.3.8.jar`) take precedence over Spark/system JARs when constructing the classloader used by `IsolatedClientLoader` on Java 8 in 'builtin' mode, which can cause unexpected behavior and/or breakages. This violates the expectation that, unless user-first classloader mode is used, Spark JARs should be prioritized over user JARs. It also seems that this separate classloader was unnecessary from the start, since the intent of 'builtin' mode is to use the JARs already existing on the regular classloader (as alluded to [here](#24057 (comment))). The isolated clientloader was originally added in #5876 in 2015. This bit in the PR description is the only mention of the behavior for "builtin": > attempt to discover the jars that were used to load Spark SQL and use those. This option is only valid when using the execution version of Hive. I can't follow the logic here; the user classloader clearly has all of the necessary Hive JARs, since that's where we're getting the JAR URLs from, so we could just use that directly instead of grabbing the URLs. When this was initially added, it only used the JARs from the user classloader, not any of its parents, which I suspect was the motivating factor (to try to avoid more Spark classes being duplicated inside of the isolated classloader, I guess). But that was changed a month later anyway in #6435 / #6459, so I think this may have basically been deadcode from the start. It has also caused at least one issue over the years, e.g. SPARK-21428, which disables the new-classloader behavior in the case of running inside of a CLI session. ### Does this PR introduce _any_ user-facing change? No, except to protect Spark itself from potentially being broken by bad user JARs. ### How was this patch tested? This includes a new unit test in `HiveUtilsSuite` which demonstrates the issue and shows that this approach resolves it. It has also been tested on a live cluster running Java 8 and Hive communication functionality continues to work as expected. Unit tests failing in #40144 have been locally tested (`HiveUtilsSuite`, `HiveSharedStateSuite`, `HiveCliSessionStateSuite`, `JsonHadoopFsRelationSuite`). Closes #40224 from xkrogen/xkrogen/SPARK-42539/hive-isolatedclientloader-builtin-user-jar-conflict-fix/take2. Authored-by: Erik Krogen <xkrogen@apache.org> Signed-off-by: Chao Sun <sunchao@apple.com>
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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 changes were proposed in this pull request?
When using the 'builtin' Hive version for the Hive metadata client, do not create a separate classloader, and rather continue to use the overall user/application classloader (regardless of Java version). This standardizes the behavior for all Java versions with that of Java 9+. See SPARK-42539 for more details on why this approach was chosen.
Why are the changes needed?
Please see a much more detailed description in SPARK-42539. The tl;dr is that user-provided JARs (such as
hive-exec-2.3.8.jar) take precedence over Spark/system JARs when constructing the classloader used byIsolatedClientLoaderon Java 8 in 'builtin' mode, which can cause unexpected behavior and/or breakages. This violates the expectation that, unless user-first classloader mode is used, Spark JARs should be prioritized over user JARs. It also seems that this separate classloader was unnecessary from the start, since the intent of 'builtin' mode is to use the JARs already existing on the regular classloader (as alluded to here). The isolated clientloader was originally added in #5876 in 2015. This bit in the PR description is the only mention of the behavior for "builtin":I can't follow the logic here; the user classloader clearly has all of the necessary Hive JARs, since that's where we're getting the JAR URLs from, so we could just use that directly instead of grabbing the URLs. When this was initially added, it only used the JARs from the user classloader, not any of its parents, which I suspect was the motivating factor (to try to avoid more Spark classes being duplicated inside of the isolated classloader, I guess). But that was changed a month later anyway in #6435 / #6459, so I think this may have basically been deadcode from the start. It has also caused at least one issue over the years, e.g. SPARK-21428, which disables the new-classloader behavior in the case of running inside of a CLI session.
Does this PR introduce any user-facing change?
No, except to protect Spark itself from potentially being broken by bad user JARs.
How was this patch tested?
This includes a new unit test in
HiveUtilsSuitewhich demonstrates the issue and shows that this approach resolves it. It has also been tested on a live cluster running Java 8 and Hive communication functionality continues to work as expected.