Key information:

• In earlier design, 40 million reads per second was achieved through deduplication but lacked end-to-end consistency.
• New implementation introduced write-through consistency protocol, deduplication layer, and tombstone markers.
• After transaction, storage engine returns commit timestamp and affected row keys to invalidate cached entries.
• Flux continues tailing MySQL binlogs and performing asynchronous cache fills.
• Engineers explained motivation as increasing demand for higher cache hit rates and stronger consistency.
• They deprecated dedicated API and enhanced telemetry and observability dashboards.

Important details:

• Cache invalidations relied on TTL and CDC before, introducing eventual consistency and delayed update visibility.
• There were issues like read-own-writes and read-own-inserts inconsistencies.
• Cache shards were reorganized for even load distribution.
• Cache Inspector tool compares binlog events to cache entries.
• TTLs for tables can be extended up to 24 hours with high cache hit rate and low latency maintained.