[HDFS-16821] Fix regression in HDFS-13522 that enables observer reads by default.

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.4.0
• Fix Version/s: 3.4.0
• Component/s: hdfs
• Labels: pull-request-available
• Target Version/s: 3.4.0
• Hadoop Flags: Reviewed
• Assignee: Simbarashe Dzinamarira
• Reporter: Simbarashe Dzinamarira
• Created: 26/Oct/22 04:40
• Updated: 28/Jan/24 07:57
• Resolved: 01/Feb/23 05:59

Description

Serving reads consistently from Observer Namenodes is a feature that was introduced in HDFS-12943.

Clients opt-into this feature by configuring the ObserverReadProxyProvider. It is important that the opt-in is explicit because for third-party reads to remain consistent, these clients then need to perform an msync before reads.

In HDFS-13522, the ClientGSIContext is implicitly added to the DFSClient thus enabling Observer reads for all clients by default. This breaks consistency guarantees for clients that haven't opted into observer reads.

https://github.com/apache/hadoop/pull/4883/files#diff-a627e2c1f3e68235520d3c28092f4ae8a41aa4557cc530e4e6862c318be7e898R352-R354

We need to return to the old behavior of only using the ClientGSIContext when users have explicitly opted into Observer reads.

Issue Links

• is related to: HDFS-13522 HDFS-13522: Add federated nameservices states to client protocol and propagate it between routers and clients.
• is related to: HDFS-16827 [RBF SBN] RouterStateIdContext shouldn't update the ResponseState if client doesn't use ObserverReadProxyProvider
• links to: GitHub Pull Request #5078