Skip to content

feat: a new workflow for epp+keda baseline vs FMA with keda (warm/hot)#1629

Open
aavarghese wants to merge 1 commit into
mainfrom
fma_keda
Open

feat: a new workflow for epp+keda baseline vs FMA with keda (warm/hot)#1629
aavarghese wants to merge 1 commit into
mainfrom
fma_keda

Conversation

@aavarghese

@aavarghese aavarghese commented Jul 12, 2026

Copy link
Copy Markdown
Contributor

Adds a new OpenShift nightly (ci-nightly-benchmark-ocp-fma-keda.yaml) that benchmarks the direct EPP+KEDA saturation autoscaling path (no WVA controller):

  1. Baseline (EPP+KEDA, no FMA) — cicd/ocp-keda
  2. FMA warm-start with EPP+KEDA — cicd/ocp-keda-fma-warmstart
  3. FMA hot-start with EPP+KEDA — cicd/ocp-keda-fma-hotstart

KEDA scales the decode/FMA-requester Deployment on EPP pool metrics (inference_pool_average_kv_cache_utilization > 0.7, inference_pool_average_queue_size > 2) via a ScaledObject → HPA, enabled purely through eppKedaSaturation.enabled: true — no VariantAutoscaling, no prometheus-adapter.

Shared observability:

  • New T_scale_up metric — autoscaler scale-up latency = HPA status.lastScaleTime → newest replica Ready. Replaces the old WVA-controller-log-derived "cold-start" row (which didn't exist on the EPP-KEDA path and is deprecated now that WVA scales via KEDA). Scaler-agnostic (reads resources.yaml), so it's identical for both nightlies. Sits alongside T_actuation (container start → Ready).
  • step_09a capture step:
    • Captures KEDA CRs (scaledobjects.keda.sh, triggerauthentications.keda.sh) in resources.yaml, plus scaledobject.txt and scaledobject-describe.txt
    • Adds EPP pool trigger-metric Thanos queries (kv-cache utilization, queue size)
    • Removes deprecated prometheus-adapter-era captures (external-metric query/discovery, prometheus-adapter logs) and the VariantAutoscaling CR
  • Display dumps: per-arm Dump ScaledObject + Dump ScaledObject describe (both nightlies); KEDA workflow drops the dead WVA-controller-log dump, WVA workflow keeps it.

🤖 Generated with Claude Code (https://claude.com/claude-code)

@aavarghese
aavarghese force-pushed the fma_keda branch 2 times, most recently from 0e28dfb to 672e064 Compare July 14, 2026 19:25
@aavarghese
aavarghese marked this pull request as ready for review July 14, 2026 19:41
@aavarghese
aavarghese requested a review from dumb0002 July 14, 2026 23:48
@aavarghese
aavarghese force-pushed the fma_keda branch 2 times, most recently from a89dd40 to 54c4fa8 Compare July 15, 2026 18:25
@Vezio

Vezio commented Jul 20, 2026

Copy link
Copy Markdown
Collaborator

@aavarghese what is the status of this - tested etc?

@aavarghese

Copy link
Copy Markdown
Contributor Author

@Vezio this one's testing is in-progress. We are not seeing the numbers we expect for hot start with FMA so investigation ongoing with @dumb0002 - kinda related to PR #1657

@aavarghese
aavarghese force-pushed the fma_keda branch 8 times, most recently from 9e326fe to f83e739 Compare July 24, 2026 20:57
@aavarghese
aavarghese enabled auto-merge (squash) July 24, 2026 20:59
@aavarghese

Copy link
Copy Markdown
Contributor Author

/cc @Vezio @dumb0002 ready for review please

Signed-off-by: aavarghese <avarghese@us.ibm.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants