TTL expiration behavior with merge operators in RocksDB compaction

Direct Answer

In standard RocksDB compaction, the TTL compaction filter runs before the merge operator. Expired keys are dropped from the input stream prior to merge-operator resolution, so the merge operator never sees a TTL-expired base value during compaction. This behavior is consistent across compaction styles (level, universal, FIFO) in RocksDB 6.0+.

Compaction Pipeline Order

1. Iterator seeks to the key range being compacted.
2. TTL compaction filter evaluates each entry's timestamp against the column family's TTL. Expired entries are marked for deletion and excluded from the output.
3. Merge operator is invoked only for surviving keys that have pending merge operands. The base value passed to FullMerge or PartialMerge is the latest non-expired value.
4. Output is written to the new SST file.

Because the filter precedes the merge, write amplification is reduced: expired keys are not rewritten, and no merge computation occurs for them. Memtable usage is unaffected during compaction; TTL expiration in memtables occurs at flush time via the same filter logic.

Caveats and Edge Cases

• Read-path (lazy) expiration: A Get() or iterator may encounter an expired key before compaction runs. The merge operator can be invoked during read-path merge resolution if the key has pending operands in the memtable. This is a separate code path from compaction.
• Custom compaction filters: If you install a custom CompactionFilter or CompactionFilterFactory that does not delegate to the built-in TTL filter, the ordering guarantee is lost.
• Merge operands with TTL: Individual merge operands written with a TTL (via WriteOptions::ttl) are subject to the same filter; they expire independently of the base key.

Configuration Options

No special configuration is required to enforce the filter-before-merge order. The following settings influence when compaction runs, not the internal ordering:

• compaction_style (level, universal, FIFO) — affects scheduling, not pipeline order.
• ttl column family option — enables the built-in filter.
• periodic_compaction_seconds — triggers compaction on idle data, reducing the window where read-path expiration applies.

Verification Step

Enable INFO level logging for the compaction and merge_operator categories, then run a test: insert a key with a 1-second TTL, wait for expiration, issue a Merge, force compaction, and confirm the log shows CompactionFilter::Filter dropping the key before any MergeOperator::FullMerge call. This single diagnostic confirms the ordering in your build.