v2.3.3

Live session state moved out of Postgres

• The four per-session live fields (hashRate, currentDifficulty, channelCount, bestDifficulty) were rewritten ~3×/minute per active session — roughly 2 000 row writes a minute at current pool size. They now live in per-session client:live:* Redis hashes with a TTL on the dead-session-sweep clock, and client_entity keeps only the birth row.
• The two bulk writers behind that load are gone, so CLIENT_ENTITY_BULK_WRITE_LOCK no longer serialises two long statements over the same ~700 rows. It still guards row births and stays for the next multi-row writer.
• Scope, measured rather than assumed: of 42 slow-statement burst minutes on the previous release, 23 contain only statistics inserts and no client_entity write at all. This removes one of two independent causes, not both.

The public high-score list survives a miner's own reset

• /api/info → highScores read the same column /bestdiff_reset zeroes, so a miner clearing their own best also erased their leaderboard entry — permanently, because the flush's GREATEST only ever re-offers the current window's maximum.
• The value is now split. bestDifficulty stays the miner's own, resettable number; allTimeBestDifficulty is the pool's record and is never lowered by any reset or delete path. Both are folded by the same upsert, so there is no second writer to drift.
• A reset now also clears the per-session bests in Redis. It never did, so a miner who reset kept seeing the old value on every worker row until the next share.
• delete-all empties the address-settings row instead of deleting it, so the record survives. What stays behind carries no address — only a difficulty, a firmware string and a timestamp.

Fixes

• The /bestdiff_reset bot command and the reset endpoint left different state behind: the endpoint deleted the notification baseline row, the bot command did not. Both now go through one function.

Upgrade notes

• Migrations 0013 and 0014 run at boot. 0013 drops four client_entity columns, and the previous image names currentDifficulty in its row-birth INSERT — an old core container can no longer create a session once the migration has run. Deploy everything at once (docker compose --profile mainnet up -d without service names), not service by service.
• Rolling back is not "put the old image back": the dropped columns have to be added again first, or the old code writes into columns that no longer exist.
• One large crons.kill_dead_clients: reconciled sessions swept=N about six to seven minutes after the deploy is expected — rows that existed before the deploy never received a client:live:* key.
• Migration 0014 seeds allTimeBestDifficulty from whatever bestDifficulty holds at that moment. An address whose best was reset shortly before the deploy carries the reset value into the leaderboard, not its historic record.