S3Guard uses a trailing log of metadata changes made to an S3 bucket to add consistency to the eventually-consistent AWS S3 service. We should add some metrics that are incremented when we detect inconsistency (eventual consistency) in S3.
I'm thinking at least two counters: (1) getFileStatus() (HEAD) inconsistency detected, and (2) listing inconsistency detected. We may want to further separate into categories (present / not present etc.)
This is related to Auth. Mode and TTL work that is ongoing, so let me outline how I think this should all evolve:
This should happen after
HADOOP-15621 (TTL for dynamo MetadataStore), since that will change when we query both S3 and the MetadataStore (e.g. Dynamo) for metadata. There I suggest that:
- Prune time is different than TTL. Prune time: "how long until inconsistency is no longer a problem" . TTL time "how long a MetadataStore entry is considered authoritative before refresh"
- Prune expired: delete entries (when hadoop CLI prune command is run). TTL Expired: entries become non-authoritative.
- Prune implemented in each MetadataStore, but TTL filtering happens in S3A.
Once we have this, S3A will be consulting both S3 and MetadataStore depending on configuration and/or age of the entry in the MetadataStore. Today HEAD/getFileStatus() is always short-circuit (skips S3 if MetadataStore returns results). I think S3A should consult both when TTL is stale, and invoke a callback on inconsistency that increments the new metrics. For listing, we already are comparing both sources of truth (except when S3A auth mode is on and a directory is marked authoritative in MS), so it would be pretty simple to invoke a callback on inconsistency and bump some metrics.
Comments / suggestions / questions welcomed.