Cache analysis and tuning

Answers "are the caches sized right and are they helping?" from system.metrics, system.asynchronous_metrics, system.events, system.parts, system.server_settings and system.metric_log. Run altinity-expert-clickhouse-connection first if the connection mode, cluster and time window are not yet established.

Query packs

• checks.sql — 12 checks: mark cache health (hit ratio and RAM share), uncompressed cache health, query cache, compiled expression cache, mark bytes per table, primary-key memory per table, cache events over time, current cache server settings, mark-cache sizing analysis, a hit-ratio diagnostic over the tables queried today, tables holding excessive marks, and the mark-cache hit ratio minute by minute for the last hour. Check caches-07 needs system.asynchronous_metric_log and caches-12 needs system.metric_log; skip them when those tables are absent.
• reference.md — background (settings, sizing, anti-patterns); read only when you need to explain a recommendation.

How to run the query packs

1. Read each pack file from this skill's directory (the skill loader prints the directory path).
2. Run statements one at a time, never a whole file. Statements end with ; and start with a -- @check <id> <title> header; keep the id with its result.
3. Honor -- @requires: skip the statement when the named table is missing, when keeper is required and the server has no Keeper/ZooKeeper, or when the version condition is not met. List skipped ids with the reason.
4. Keep {cluster} as written when a cluster macro exists; otherwise apply the connection skill's rewrite rule. Any other {placeholder} is a template variable: substitute a real value first or skip the statement.
5. On an error, record the check id and the first line of the error, then continue. Only for UNKNOWN_IDENTIFIER, run DESCRIBE TABLE system.<table> and drop the missing column.
6. A severity column is the verdict for that row. Copy it; do not re-grade.