We have run performance evaluations on the version of push-based shuffle committed to upstream so far, and have identified a few places for further improvements:
1. On the server side, we have noticed that the usage of `String.format`, especially when receiving a block push request, has a much higher overhead compared with string concatenation.
2. On the server side, the usage of `Throwables.getStackTraceAsString` in the `ErrorHandler.shouldRetryError` and `ErrorHandler.shouldLogError` has generated quite some overhead.
These 2 issues are related to how we are currently handling certain common block push failures.
We are communicating such failures via `RPCFailure` by transmitting the exception stack trace.
This generates the overhead on both server and client side for creating these exceptions and makes checking the type of failures fragile and inefficient with string matching of exception stack trace.
To address these, this PR also proposes to encode the common block push failure as an error code and send that back to the client with a proper RPC message.
Improve shuffle service efficiency for push-based shuffle.
Improve code robustness for handling block push failures.
No
Existing unit tests.
Closes#33613 from Victsm/SPARK-36378.
Lead-authored-by: Min Shen <mshen@linkedin.com>
Co-authored-by: Min Shen <victor.nju@gmail.com>
Signed-off-by: Mridul Muralidharan <mridul<at>gmail.com>
(cherry picked from commit 3f09093a21)
Signed-off-by: Mridul Muralidharan <mridulatgmail.com>
c6b683e5a2