Andrew Kyle Purtell created PHOENIX-8056:
--------------------------------------------
Summary: Top-N queries incur an unnecessary round-trip per region
Key: PHOENIX-8056
URL: https://issues.apache.org/jira/browse/PHOENIX-8056
Project: Phoenix
Issue Type: Bug
Components: core
Affects Versions: 5.3.2
Reporter: Andrew Kyle Purtell
Assignee: Andrew Kyle Purtell
Fix For: 5.4.0, 5.3.3
Global index scans executing Top-N queries can incur an unnecessary round-trip
per region following PHOENIX-7707. During scanner initialization, range scans
using the default read type that exceed {{hbase.storescanner.pread.max.bytes}}
trigger HBase's {{returnImmediately}} flag (HBASE-20457) on the open scanner
context.
An extra RPC returns a buffered dummy tuple without advancing the delegate
scanner. For uncovered index scans, this inflated communication costs to three
RPCs per region where two should suffice.
To eliminate the redundant network hop, the scanner lifecycle adopts an eager
"pull-first" strategy on all invocations after the initial call whenever a
dummy tuple is held.
This change also corrects tenant memory tracking during deferred aggregation.
Previously, the Top-N memory chunk size was determined from the iterator's byte
size after fetching the first tuple. When that first tuple was an incomplete
dummy due to early pagination, the byte size registered as zero and was never
subsequently adjusted. As a result, scans that paged out during scanner open
held a 0-byte memory reservation for the entire Top-N heap.
The the fix is to recalculate and resize the allocation chunk once when the
build actually concludes, either upon consuming the first real tuple or upon
exhausting the scan.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)