Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
3.3.1
Description
Bug description:
The read method in S3AInputStream has this following behaviour when an IOException happening during the read:
- reopen and read quickly: The client after failing in the first attempt of read, will reopen the stream and try reading again without sleep.
- reopen and wait for fixed duration: The client after failing in the attempt of read, will reopen the stream, sleep for fs.s3a.retry.interval milliseconds (defaults to 500 ms), and then try reading from the stream.
While doing the reopen and read quickly process, the subsequent read will be retried without reopening the input stream in case of the second failure happened. This leads to some of the bytes read being skipped which results to corrupt/less data than required.
Scenario to reproduce:
- Execute S3AInputStream `read()` or `read(b, off, len)`.
- The read failed and throws `Connection Reset` exception after reading some data.
- The InputStream is re-opened and another `read()` or `read(b, off, len)` is executed
- The read failed for the second time and throws `Connection Reset` exception after reading some data.
- The InputStream is not re-opened and another `read()` or `read(b, off, len)` is executed after sleep
- The read succeed, but it skips the first few bytes that has already been read on the second failure.
Proposed fix:
https://github.com/apache/hadoop/pull/3109
Added the test that reproduces the issue along with the fix
Attachments
Issue Links
- is depended upon by
-
HADOOP-17812 NPE in S3AInputStream read() after failure to reconnect to store
-
- Resolved
-
- relates to
-
HADOOP-15541 AWS SDK can mistake stream timeouts for EOF and throw SdkClientExceptions
-
- Resolved
-
- links to