linuser/zabbix-network-topology v5.6.11

v5.6.11 — single-flight cold rebuilds, serve last-good snapshot (#35 part 1)

on GitHub

Fixed

• Concurrent cold rebuilds no longer pile up — simultaneous refreshes share one rebuild and get the last good snapshot meanwhile (#35, part 1). The network.topology.data request caches its expensive payload for 60 s, but a miss (expired, first open, a tab returning after a pause) rebuilt the whole graph, and several of those at once each ran the full build. On a large, partitioned MariaDB history the items:last fallback can scan old partitions long enough to hit the gateway's read timeout (HTTP 504), and the pile-up made it worse.

  Now, per user and map selection, a single-flight lock lets exactly one request rebuild; concurrent requests serve the last good snapshot (a longer-lived copy of the last successful build) instead of launching their own rebuild, and a genuinely cold request with no snapshot yet waits briefly for the in-flight build rather than duplicating it. The lock self-heals via a TTL, and without APCu everything is fail-open — every request builds exactly as before.

  This addresses the "simultaneous refreshes should share the rebuild" half of #35; shortening the single cold build itself (prepared history from closed partitions) is tracked separately. Reported with detailed measurements by @rhysjoseph.