Autoscale Profile

An autoscale profile is a named preset that expands into a complete operatorBox.autoscale block at Katalog load time.

Profiles are relative — they use the CRD’s declared workers and queue.maxQueueDepth as a baseline, then compute worker overrides, queue overrides, trigger thresholds, and timing from that baseline.

Profiles

Profile	Trigger	Workers override	Queue override	Interval	Cooldown
`burst`	queueDepth > 60% of override queueDepth	baseline × 4	baseline × 10	5s	30s
`steady`	queueDepth > 40% AND workersBusy > 70%	baseline × 2	baseline × 3	30s	2m
`batch`	cron window 23:00 → 02:00	baseline × 3	baseline × 8	60s	5m
`latency-sensitive`	P95 reconcile > 200ms	⌈baseline × 2.5⌉	—	15s	1m
`cost-optimized`	workersIdle > 60% AND queueDepth > 80%	max(1, baseline × 0.5)	baseline × 0.5	30s	10m

Usage

operatorBox:
  workers: 4
  queue:
    maxQueueDepth: 100
  autoscale:
    profile: steady

For steady with the above baseline (workers=4, queueDepth=100):

override workers = 4 × 2 = 8
override queueDepth = 100 × 3 = 300
trigger threshold = 300 × 40% = 120
trigger condition: queueDepth > 120 AND workersBusyPercent > 70

Rules

Profile or explicit — not both.

# Valid
autoscale:
  profile: burst

# Valid
autoscale:
  interval: 10s
  cooldown: 1m
  conditions:
    when:
      - field: metrics.queueDepth
        greaterThan: "500"
  do:
    workers: 16

# Invalid — rejected at load time
autoscale:
  profile: burst
  interval: 10s

Unknown profiles fail fast. An unrecognized name is a Katalog load error.

Profile details

`burst`

React instantly to spikes. Aggressive scaling, short interval, short cooldown.

trigger:  queueDepth > 60% of override queueDepth
override: workers × 4, queueDepth × 10
timing:   interval 5s, cooldown 30s

Use when your operator handles unpredictable bursts and you need it to scale out before the queue fills.

`steady`

Smooth, predictable scaling. Requires both queue pressure and worker saturation before scaling.

trigger:  queueDepth > 40% of override queueDepth AND workersBusyPercent > 70
override: workers × 2, queueDepth × 3
timing:   interval 30s, cooldown 2m

Use for production operators with consistent load patterns.

`batch`

Scale for a nightly processing window. Time-triggered, not load-triggered.

trigger:  cron "0 23 * * *", duration 3h
override: workers × 3, queueDepth × 8
timing:   interval 60s, cooldown 5m

Use when you have a known batch window (nightly jobs, end-of-day processing).

`latency-sensitive`

Keep reconcile latency low. Triggered by P95 latency, not queue depth.

trigger:  reconcileDurationP95Ms > 200
override: workers × 2.5 (ceiling)
timing:   interval 15s, cooldown 1m

Use when users are sensitive to reconcile delay and you want to scale out before latency compounds.

`cost-optimized`

Minimize resource usage during low activity. Scale down aggressively when idle.

trigger:  workersIdlePercent > 60 AND queueDepth > 80% of override queueDepth
override: max(1, workers × 0.5), queueDepth × 0.5
timing:   interval 30s, cooldown 10m

Use for operators with long idle periods between bursts, where you want to give back resources between processing windows.

Choosing a profile

Situation	Profile
Unpredictable spikes, must not queue-block	`burst`
Consistent throughput, standard production	`steady`
Known batch window (nightly, weekly)	`batch`
Latency SLO, users feel slow reconciles	`latency-sensitive`
Long idle periods, cost matters	`cost-optimized`

If none fit, write the autoscale block manually.