Pull Request Template

Description

This change contains the fix for occasional hang that can occur while querying using partition partition key against a sub-partitioned container. Let's assume a container with partition key definition "/id", "/value1", "/value2". Unlike single-key partition definition, for hierarchical partitions, partition splits occurs such that for a certain value of "/id" (say 2), more than one partition contains the data for this (partial) partition value. When a query is issued with partial partition key ("/id" = 2 in this case), query enumerator (QueryPartitionRangePageAsyncEnumerator L50) attempts to query a feedrange based on this partial value that isn't fully contained inside a single physical partition in the backend. Since such a range cannot be resolved to any single physical partition, enumerator receives a GoneException from the underlying stack. The GoneException is caught by upstream (CrossPartitionRangePageAsyncEnumerator L114) which attempts to resolve the physical partition ranges by consulting the routing map. This is typically a no-op since the GoneException isn't received because of a change in state of the backend.
This doesn't actually help since CrossPartitionRangePageAsyncEnumerator attempts to locate FeedRangeEpk (without the partial partition key, which honors partition boundary) inside a physical partition range - which always succeeds and further downstream (RequestInvokerHandler.ResolveFeedRangeBasedOnPrefixContainerAsync and RequestInvokerHandler L244) ignores this FeedRangeEpk and instead queries using partial partition key (FeedRangePartitionKey) which spans across the partition boundary. The later continues to cause another GoneException and cycle repeats.

This fix limits the epk range for QueryPartitionRangePageAsyncEnumerator such that it does not span across the single physical partition range. It also uses FeedRangeEpk to represent this new partition range, which causes downstream to no longer perform extraneous operations (such as ResolveFeedRangeBasedOnPrefixContainerAsync).

Several parts of the code are suspect here and a more comprehensive fix may be needed to handle all cases generally, outside this change.

1. RequestInvokerHandler.ResolveFeedRangeBasedOnPrefixContainerAsync should not happen. Ignoring the FeedRange supplied by the upstream blindly is plain wrong. Instead upstream enumerators etc. should provide explicit range.
2. At the very least using type of FeedRangeInternal (Epk v/s Partition) to determine extra processing in RequestInvokerHandler should not exist. This should be an explicit action.
3. It's not clear how filtering occurs in the backend when partial partition key is specified and request lands on a physical partition where data for other logical partitions also exists. We need to replicate this behavior in the test infrastructure in the SDK so that such scenarios can be tested broadly.
4. Query enumerators should not take a dependency on partition key. Instead partition key -> physical partition resolution should take place upstream before they are instantiated.

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• [] New feature (non-breaking change which adds functionality)
• [] Breaking change (fix or feature that would cause existing functionality to not work as expected)
• [] This change requires a documentation update

Closing issues

To automatically close an issue: closes #4326