PromQL offset / @ time-shift modifiers have no IR representation

[milindsrivastava1997]

Problem

PromQL's offset modifier (metric offset 1h) and @ modifier (metric @ 1609746000) shift the evaluation/lookback time of a vector selector. Unlike without(...) (#39), which is blocked on missing catalog metadata, this is blocked on a missing IR node: there is no time-shift concept anywhere in relational::QueryExpr or canonical QueryExpr to carry the shift amount, so both modifiers are rejected unconditionally rather than partially/incorrectly lowered.

These are common in real dashboards and alerts (week-over-week comparisons, anomaly detection baselines), so this is a meaningful coverage gap distinct from the scalar-operand and without gaps.

Reproduction

crates/lower/src/promql.rs, vs_parts:

fn vs_parts(vs: &VectorSelector) -> Result<(String, Vec<L2Expr>)> {
    // `offset` / `@` shift the evaluation/lookback time. The intent algebra has
    // no representation for either, so silently lowering them (as if absent)
    // would change the query's meaning. Reject rather than mislower.
    if vs.offset.is_some() || vs.at.is_some() {
        return Err(LoweringError::UnsupportedFeature(
            "`offset` / `@` time-shift modifiers have no intent-algebra representation".into(),
        ));
    }
    ...

Any query with offset or @ on a vector/matrix selector — e.g. rate(http_requests_total[5m] offset 1h) or http_requests_total @ 1609746000 — fails to lower with LoweringError::UnsupportedFeature. No test exercises a successful lowering of either modifier.

[zzylol]

Reopening: this was closed as completed, but the fix never landed on main.

The implementation commit fff0a2e (feat(promql): lower offset / @ time-shift modifiers to a TimeShift node (#40)) exists only on the unmerged branch feat/40-offset-at-modifiers:

$ git merge-base --is-ancestor fff0a2e main; echo $?
1                      # not an ancestor
$ git branch -a --contains fff0a2e
  feat/40-offset-at-modifiers
  remotes/origin/feat/40-offset-at-modifiers

Unlike the other closed feature issues (#84, #89, …), whose branches show as unmerged only because they were squash-merged, there is no equivalent squashed commit for this one on main, and grep -rn TimeShift crates/ir/src/ returns nothing.

crates/frontend-promql/src/promql.rs:1414 on main still rejects unconditionally:

http_requests_total offset 5m
→ unsupported feature: `offset` / `@` time-shift modifiers have no intent-algebra representation

Corpus impact (measured on main)

65 rejections in the Prometheus testdata corpus, and it is the only failure mode in the official-docs corpus — all 4 failing docs queries are offset/@:

http_requests_total offset 5m
sum(http_requests_total{method="GET"} offset 5m)
http_requests_total @ 1609746000
http_requests_total @ 1609746000 offset 5m

That makes it the highest-impact real-world PromQL gap remaining. Either feat/40-offset-at-modifiers needs to be rebased and merged, or this issue should be reclosed as not planned with the branch's rationale recorded.

[zzylol]

Correction to the corpus figures above: the docs corpus has 5 failures, not 4. Four are offset/@; the fifth is {__name__=~"job:.*"}, which is rejected deliberately per #67.

Verified output:

docs corpus: Tally { lowered: 44, rejected: 5, unparseable: 0 }

[1] {__name__=~"job:.*"}                              => non-equality `__name__` matcher (per #67)
[2] http_requests_total offset 5m                     => offset / @
[3] sum(http_requests_total{method="GET"} offset 5m)  => offset / @
[4] http_requests_total @ 1609746000                  => offset / @
[5] http_requests_total @ 1609746000 offset 5m        => offset / @

The substance is unchanged: offset/@ is 4 of the 5 docs-corpus failures and the only unintentional one, plus 65 rejections in the testdata corpus.

[zzylol]

Retraction — I was wrong, re-closing.

offset / @ is supported on main. This issue was correctly closed, and my reopen was based on a stale local ref. Apologies for the noise.

What happened: my local main was 2 commits behind origin/main, and the branch I was measuring from (test/asapquery-sql-corpus) is based on that older main. So git merge-base --is-ancestor fff0a2e main was answering about a main that predated the merge. PR #106 merged feat/40-offset-at-modifiers into main on 2026-07-06 (merge commit 7f907a98ba), and origin/main does contain the work:

$ git show origin/main:crates/ir/src/intent_algebra/query_expr.rs | grep -c TimeShift
16
$ git show origin/main:crates/frontend-promql/src/promql.rs | grep -n TimeShift | head -1
47://! | `<selector> offset <d>` / `<selector> @ <ts>`/`start()`/`end()` | `TimeShift{shift}` …

Verified by running the corpus against a clean origin/main worktree (9fdd3d7):

corpus	on stale branch	on origin/main
docs	44 lowered, 5 rejected	48 lowered, 1 rejected
testdata	1441 lowered, 147 rejected	1502 lowered, 86 rejected

The 65 offset/@ rejections are gone. The single remaining docs rejection is {__name__=~"job:.*"}, which is the deliberate #67 behaviour.

Both of my earlier comments on this issue should be disregarded. I've corrected the corpus figures in #108, #109 and #112, which were all measured against the same stale tree — the parser gaps in #108 are unaffected (still exactly 235 parse failures), but the denominators changed.