Summary

This PR provides a definitive fix for a bug in the PGWire INSERT cache that could cause intermittent NullPointerException errors. It resolves the underlying issue that was only partially mitigated before.

Symptoms

When executing SQL INSERT statements via the PostgreSQL wire protocol, the server would occasionally fail with one of the following errors in the logs:

Internal error. Exception type: NullPointerException

or

Cannot invoke "io.questdb.cairo.sql.InsertOperation.createMethod(io.questdb.griffin.SqlExecutionContext, io.questdb.cairo.pool.WriterSource)" because "this.insertOp" is null

Affected Versions

This issue is present in QuestDB versions from 8.2 up to and including 8.3.3.

Context & Root Cause

While PR #5706 prevented these specific errors from appearing, it did not fix the core problem: the INSERT cache was creating duplicate entries and when one entry was closed on eviction, it also closed the other entry which was left in a cache and on usage it would throw the internal error. This PR corrects the cache logic, fully resolving the bug.

Implementation details

The put() method was creating a large number of duplicate entries in INSERT caches used by the PGWire protocol. This was triggered when PGWire would peek() at a cached entry and later re-put() it; The old logic would fail to find the existing entry unless it was at the head of its block, resulting in a duplicate.

This commit changes put() to scan the entire block for a matching key. The new behavior is:

1. If an identical (key, value) pair exists, it is promoted to the MRU position.
2. If a (key, null) entry exists (left by poll()), its slot is reused for the new value to avoid a key allocation.
3. Otherwise, a new entry is inserted.

This change introduces a loop to scan the block, which adds a small seek cost. However, given the small block sizes used in practice (e.g., the default of 4 for the INSERT cache), this cost is negligible and is an acceptable trade-off for correctness and the elimination of duplicate entries.