Sizing memory and pollers

One instance runs one poller per dashboard and set of viewer picks (Scaling). Memory is what bounds how many pollers it can hold, and this page is the arithmetic.

What a poller holds

A poller keeps each panel’s last result, so a viewer who joins is shown the dashboard at once rather than after the next cycle. That is most of its heap. Measured on a 20-panel dashboard of time series (#495):

Rows per panel	Heap per poller
720	1.6 MiB
5,000 (MAX_QUERY_ROWS)	10.7 MiB

Viewers add little on top of that. A cycle is serialized once for everyone on the same cursor, and each stream buffers at most 1 MiB before it counts as behind (Streaming).

The heap and the cap

• The heap ceiling. The chart sets NODE_OPTIONS=--max-old-space-size to three quarters of resources.limits.memory, so under pressure V8 collects garbage instead of the kernel killing the pod. A limit below 256Mi fails the render. The image’s own default, for docker run, is 768 MB, and the quick start’s is 512 MB, because that container also runs Postgres.
• The poller cap. MAX_POLLERS is the most pollers one process runs. Left at 0, it is half the heap ceiling over 11 MB, the worst case above: 34 pollers at the chart’s default 1Gi limit. A stream that would need one more poller is refused. The browser is told the server is at its limit and reconnects by itself a few seconds later, and other callers get 503 with Retry-After. Viewers of a running poller are never affected.

Two metrics show the headroom: holotable_pollers_active against holotable_pollers_limit (Metrics).

Worked example

Take 50 dashboards of 20 panels at 720 rows each, every one with a different time range or variable pick, so none share a poller. That needs 50 × 1.6 MiB, which is 80 MiB of results, comfortably inside a 1Gi pod. At the 5,000-row cap the same dashboards need 535 MiB. That is past the derived cap of 34 pollers, so the 35th distinct dashboard waits. Raise resources.limits.memory, which raises the heap and the derived cap with it, or set MAX_POLLERS explicitly when the dashboards are known to be small.