[DRAFT] Add physicsShedder filter for latency-aware load shedding#3973
Draft
ivan-digital wants to merge 4 commits into
Draft
[DRAFT] Add physicsShedder filter for latency-aware load shedding#3973ivan-digital wants to merge 4 commits into
ivan-digital wants to merge 4 commits into
Conversation
Introduces physicsShedder, a route-level filter that catches gray failures (slow-but-200 backends) by modeling incoming traffic as resistance against a learned baseline. R = avgLatency/latencyTarget + errorWeight * errorRate threshold = mu + k*sigma (EWMA baseline + adaptive deviation) pReject = max(0, (R - threshold) / R), clamped to 0.95 Complements admissionControl, which reacts only to error rate. Uses the same Admission-Control header convention so the two filters compose on one route without double-counting. Known v1 tradeoff: a sudden step in latency inflates the adaptive variance, which can keep the threshold above R during the transient and let the new latency become the baseline. Collect data in logInactive or inactive mode first; refine threshold formulation based on feedback. Closes zalando#3828 Tests include unit math, warmup gate, ring buffer rotation, mode behavior, pre/post processor, response error counting, metrics emission, tracing spans, fuzz on the math, randomized invariants, concurrent hot path, admissionControl chain composition, and a local load-test script under skptesting/. Signed-off-by: ivan-digital <root@ivan.digital>
ivan-digital
changed the title
added 3 commits
April 18, 2026 13:30
Signed-off-by: ivan-digital <root@ivan.digital>
Signed-off-by: ivan-digital <root@ivan.digital>
Signed-off-by: ivan-digital <root@ivan.digital>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
New filter
physicsShedderthat watches latency to catch slow-but-200 backends (gray failures) — the caseadmissionControlmisses because it only looks at error rate.It learns what "normal" latency looks like for a route and starts rejecting some traffic when latency clearly drifts above that. Composes with
admissionControlon the same route — both honor theAdmission-Controlresponse header so they don't double-count.Closes #3828.
Known tradeoff
A sudden latency step (e.g. backend goes 50ms → 100ms) doesn't trigger shedding right away, because the noise estimate spikes with the change and ends up "explaining away" the spike. Confirmed with a local load test against a slow backend.
The filter still catches error bursts and gradual degradation. For the sudden-step case, would love feedback on the right fix — capping the threshold, lowering the multiplier, or bringing back a memory term.
Tests
Unit, fuzz, randomized invariants, race-checked, scenario tests, tracing, composition with
admissionControl, plus a local load script underskptesting/.go test ./... -raceclean. Verified the metrics endpoint emits the expected names against a running skipper.