mirror of
https://github.com/milvus-io/milvus.git
synced 2026-10-29 01:16:06 +08:00
zhenshan.caoandClaude Opus 5
b67818e600
Cherry-pick of #51496 pr: #51495 issue: #51493 ### Why this branch needs it #51496 landed this fix on `hotfix-2.6.20` (shipped as v2.6.20) but **never on `2.6`**. `2.6` still carries the v2.6.19 in-place status split — `metrics.go` has `FailInputLabel` / `FailSystemLabel` / `RejectedUserLabel` / `RejectedSystemLabel`, and `req_count` has no `cause` label at all. So the next release cut from `2.6` would emit `fail_input` / `fail_system` again: a user upgrading **v2.6.20 -> 2.6.21** would see the dashboards and alert rules that v2.6.20 had just repaired break a second time — and this time as a regression from a working state, which is harder to notice than the original break. ### What #50221 split the `status` label values of `milvus_proxy_req_count` / `milvus_proxy_grpc_latency` in place. It shipped in **v2.6.19**, so upgrading from <= v2.6.18 silently zeroes every existing dashboard panel and alert rule matching `status="fail"` / `status="rejected"` — an alert that stops firing rather than erroring. This keeps the classification but moves it onto its own dimension, restoring the status domain to what <= v2.6.18 emitted: ``` status: success | fail | rejected | retry | total | abandon (as before v2.6.19) cause: user | system | cancel (new, only on fail/rejected) ``` | v2.6.19 / v2.6.20 | this PR | |---|---| | `status="fail_input"` | `status="fail", cause="user"` | | `status="fail_system"` | `status="fail", cause="system"` | | `status="rejected_user"` | `status="rejected", cause="user"` | | `status="rejected_system"` | `status="rejected", cause="system"` | | `status="cancel"` | `status="fail"` or `"rejected"`, `cause="cancel"` | | `success` / `retry` / `total` / `abandon` | unchanged, **no cause** | Outcomes that have no responsible party emit `cause=""`. Prometheus treats an empty label value as equivalent to the label being absent, so those series stay byte-identical to what pre-2.6.19 emitted — the new label materializes only where it carries information. Old queries work unchanged and still count each request once: `status="fail"` is again every hard failure, because Prometheus aggregates over `cause` for free. ### Cherry-pick notes `git cherry-pick` of #51496 applied with **no conflicts** (19 files). Cross-checked file by file against `hotfix-2.6.20`: 16 files are byte-identical, and the 3 that differ do so only in `2.6`-only content unrelated to this change (`internal/proxy/util.go` — `validateElementType`; `handler_v2.go` and `util_test.go` — surrounding `2.6` code). The cause-related lines in all three were extracted and compared separately: identical. The emit surface was **re-verified on this branch rather than assumed**: all 12 `WithLabelValues` sites for the two metrics were checked for label arity by a paren-balanced scan, because adding a label makes arity a runtime panic, not a compile error. (`2.6` has 12 sites where master/3.0 have 62 — it lacks the `InterceptorImpl` / `snapshot_impl.go` surface.) No stale `FailInputLabel` / `FailSystemLabel` / `RejectedUserLabel` / `RejectedSystemLabel` references remain, and all `ParseMetricLabel` / `failMetricLabel` call sites use the two-value signature. ### Verification on this branch - Compiles: `./internal/...` and the `pkg` module. - Green: `internal/proxy`, `internal/distributed/proxy`, `internal/distributed/proxy/httpserver`, `internal/distributed/proxy/client`, `pkg/metrics`, `pkg/util/requestutil`, `pkg/util/merr`. - `make verifiers`: 0 issues. ### Related - #51496 — `hotfix-2.6.20` (merged, the version this is taken from) - #51526 — `3.0` - #51907 — `master` (aligns `CauseNA` there from `"na"` to `""`) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Signed-off-by: zhenshan.cao <zhenshan.cao@zilliz.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>