[HADOOP-16852] ABFS: Send error back to client for Read Ahead request failure

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.3.1
• Fix Version/s: 3.3.1
• Component/s: fs/azure
• Labels: pull-request-available
• Assignee: Sneha Vijayarajan
• Reporter: Sneha Vijayarajan
• Created: 11/Feb/20 17:49
• Updated: 19/Aug/21 12:41
• Resolved: 28/May/20 03:58

Description

Issue seen by a customer:

The failed requests we were seeing in the AbfsClient logging actually never made it out over the wire. We have found that there’s an issue with ADLS passthrough and the 8 read ahead threads that ADLSv2 spawns in ReadBufferManager.java. We depend on thread local storage in order to get the right JWT token and those threads do not have the right information in their thread local storage. Thus, when they pick up a task from the read ahead queue they fail by throwing an AzureCredentialNotFoundException exception in AbfsRestOperation.executeHttpOperation() where it calls client.getAccessToken(). This exception is silently swallowed by the read ahead threads in ReadBufferWorker.run(). As a result, every read ahead attempt results in a failed executeHttpOperation(), but still calls AbfsClientThrottlingIntercept.updateMetrics() and contributes to throttling (despite not making it out over the wire). After the read aheads fail, the main task thread performs the read with the right thread local storage information and succeeds, but first sleeps for up to 10 seconds due to the throttling.

Issue Links

• breaks: HADOOP-17301 ABFS: read-ahead error reporting breaks buffer management
• links to: GitHub Pull Request #1898
• links to: PR Link