Alerting

The log reports a persistent condition once, on the transition, rather than every interval, which is right for reading and useless for alerting. These are the series to build on instead, served on /metrics (default :9110; /healthz answers 200 while the process is up).

Is it working at all

fpm_tune_apply_enabled                    # 0 means this process only watches
fpm_tune_apply_blocked{reason}            # asked to apply and cannot
fpm_tune_last_run_timestamp_seconds       # not advancing means the loop has stalled
fpm_tune_last_apply_timestamp_seconds     # not advancing while changes are pending
fpm_tune_applies_failed_total
fpm_tune_rollbacks_total                  # above zero deserves a look at the log
fpm_tune_rollback_failed_total            # worse: a rejected file is still on disk
fpm_tune_repairs_total                    # it had to undo something a run left behind

apply_blocked is the one people forget. A process that is watching and one that is acting look identical from outside, and that difference is the whole question when a host is not being tuned. Its reason label is the actionable part:

no_master: nothing to apply to.
lock: another fpm-tune holds the pool directory.
unrepaired: a previous run left something the tool could not resolve.
budget_unconfirmed: php-fpm's own memory limit could not be read, so the tool refuses to write from the machine's. Pass --memory. See The budget.
state_unsaved: an apply succeeded but its record could not be written, so the reload brake is not on disk.

Is the host full

fpm_tune_pool_demand_unmet{pool}   # this pool wanted more workers than it got
fpm_tune_capacity_exhausted        # ...and that is true of at least one pool

Either one means no configuration change will help. The machine needs more RAM, or fewer sites. See telling needs-more from machine-full.

What it is sizing to

fpm_tune_pool_workers_configured{pool}
fpm_tune_pool_workers_recommended{pool}
fpm_tune_pool_worker_rss_bytes{pool,estimate}   # typical_peak, high_water, p50, p95, p99
fpm_tune_pool_baseline_confidence{pool}         # 0 to 1; below 1 the pool will not be cut
fpm_tune_pool_measured{pool}                    # 1 when sized from its own memory, not a guess
fpm_tune_budget_bytes{state}                    # total, reserved, allocated, free

The p50/p95/p99 estimates are the measured spread, which is what to graph when a host misbehaves at its busiest minute rather than on average.

Two you should not see above zero

fpm_tune_pools_unreachable   # a pool could not be scraped this round
fpm_tune_pools_ambiguous     # two masters share a pool name; those pools are NOT published

pools_ambiguous above zero means the daemon is running unscoped on a multi-master host and two pools share a name. www is the default in every distribution. Those pools are suppressed from the per-pool series rather than published under a colliding label; name a master with --drop-in-dir. See The trust boundary.