[
https://issues.apache.org/jira/browse/CASSANDRA-21462?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18099853#comment-18099853
]
Jon Haddad commented on CASSANDRA-21462:
----------------------------------------
First patch to address the current cursor path. No new types added, just fixes
for what already exists. Here's the summary:
*Data-correctness / corruption fixes:*
- Same-timestamp writes to the same cell resolved to the wrong value (tie-break
was inverted relative to the reference reconciliation rules) — silent replica
divergence
- Tombstone vs. expiring-cell tie-break was dead code (an ambiguous
"isExpiring" check meant both sides always looked the same), letting an
expiring cell wrongly win over a tombstone at the same timestamp
- Tombstone cells were serialized with a spurious "expiring" flag and a wasted
TTL field
- Reversed (DESC) clustering columns compared empty vs. valued values in the
wrong order, causing rows to merge out of order and corrupting stats bounds
- NO_DELETION_TIME live markers were misclassified as deleted, poisoning
min-timestamp/tombstone stats
- Large (64+ column) table row encoding had two separate corruption bugs: a
flipped mode-selection at the encoding boundary, and a wrong loop bound that
silently dropped present columns sorted after the last missing one
- Index corruption for partitions larger than 2GiB (an index offset overflowed
as a signed int)
- The final output index block excluded the end-of-partition marker byte, and
partitions crossing the index-block-size threshold exactly once got no promoted
index at all
- Dropped columns weren't filtered from old cell data — a column dropped and
later re-added with the same name could resurrect stale data
- Static-row presence was derived from current table metadata instead of the
actual input sstables, breaking after ALTER TABLE ... DROP of the last static
column
- Row/static-row stats were over-counted for partitions with no static values
(an empty placeholder row was counted when it shouldn't have been)
- Disabled bloom filters weren't tolerated by the index writer
- Accord-enabled tables were purging tombstones relative to the wrong time
bound instead of gcBefore
- A shared file-rewriting utility (SSTableRewriter) didn't correctly retain
original input files in one code path when asked to
*Performance (garbage-free) fixes:*
- Eliminated a per-row heap allocation when decoding sparse rows (rows missing
some columns) under 64 columns
- Eliminated the same class of allocation for sparse rows in wide (64+ column)
tables, which used a different, previously-unoptimized decode path
> Add byte for byte test harness to detect and resolve cursor deviations from
> iterator path
> -----------------------------------------------------------------------------------------
>
> Key: CASSANDRA-21462
> URL: https://issues.apache.org/jira/browse/CASSANDRA-21462
> Project: Apache Cassandra
> Issue Type: Sub-task
> Components: Local/Compaction
> Reporter: Jon Haddad
> Assignee: Jon Haddad
> Priority: High
> Fix For: 6.0
>
>
> This issue addresses various deviations and inconsistencies with the cursor
> path by introducing tests that verify both the old iterator path and the new
> cursor path generate the same outputs given the same inputs. It also fixes
> various issues found in the process of creating this test system that
> resulted in corruption or incorrect values being carried into the resulting
> SSTable.
> It adds a test that uses the JVM's internal memory tracking to assert that we
> do NOT do unnecessary allocations in the cursor path that should guard
> against future regressions, and fixes an unnecessary enum allocation.
> It does NOT add support for BTI, multi-cell columns, counters, etc.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
