[frankh]

Why

PromQL queries must return useful results when sample times do not match query steps. Short counter windows must include the previous sample, and counter increases must not estimate values at window edges.

Keep the PromQL-compatible API and add selected MetricsQL improvements for automatic query windows, counter calculations, and supported syntax extensions. Apply the same behavior to both ClickHouse storage formats. Preserve stale markers and return numeric constants as vectors.

Other functions and native histogram calculations retain Prometheus behavior. Automatic selectors and counter rollups use the evaluation engine and can read more raw samples. Existing query limits still apply.

Validation

• go mod verify, make lint, and go build ./....
• CH_SCHEMA_LAYOUT=current SNUFFLE_E2E_SKIP_DOCKER=1 make integration-test.
• CH_SCHEMA_LAYOUT=posthog SNUFFLE_E2E_SKIP_DOCKER=1 make integration-test.

Both integration runs use real ClickHouse and include the full suite with race checks. Counter regression values also match VictoriaMetrics v1.152.0. The final documentation-only update passes git diff --check.

---

Created with PostHog Desktop

[frankh]

Pushed 400c43a with fixes for the review findings.

Correctness

• metricsQLValue builds new result slices. The engine returns the input matrix to its pools on close, so the result must not alias it.
• Prometheus functions that MetricsQL does not define (histogram_count, histogram_sum, info, ...) work again. Their names travel through the MetricsQL parser as a string argument of union.
• timestamp(m) and absent(m) keep their selector argument, so the engine returns the sample timestamp and the absent labels.
• The rollup argument index comes from metricsql.GetRollupArgIdx.
• Omitted windows and step units inside a subquery use the subquery step.
• rate returns no value when the sample duration is zero.
• Query stats are recorded after the NaN filter.

Fast paths

• Instant queries: the fast path unwraps the internal default_rollup to its selector, because both read the latest non-stale sample in the lookback window. sum, topk, and nested counts over bare selectors use SQL again.
• Range queries: automatic selector windows depend on each series and counter rollups use MetricsQL math, so SQL cannot reproduce them. The unreachable range SQL paths and their tests are removed. metricsQLServer and the per-request queryable copy are gone; no read path uses REMOTE_WRITE_SAMPLE_INTERVAL now.

running_sum is a post-pass over the range result. It no longer doubles the scanned range and no longer requires start aligned to step. It must be the outermost function.

Docs: README and PERFORMANCE describe the paths that run, including the one-step visibility of a single sample in a range query.

Validation here: make lint, go build ./..., and go test -race ./... pass. The ClickHouse integration suite was not run in this environment. The e2e table gained timestamp and absent cases and a rangePoints column; please run make integration-test for both layouts.