trickstercache / trickster / 27907115766

89%

Build Type

push

github

Committed by web-flow

Commit Message

fix+test: [timeseries] panic on ExtentList.Splice when shard count exceeds len(el)*4 hint (#1016)

* fix+test: [timeseries] panic on ExtentList.Splice when shard count exceeds len(el)*4 hint

spliceByTime and spliceByTimeAligned pre-allocated their result slice
with a fixed `len(el)*4` length and used `out[k] = ...` indexed writes.
When the requested time range produces more than 4 shards per input
extent — easily reached on any backend running with shard_max_size_time
(e.g., 6h range / 1h shard = 6 shards, 30d / 1h = 720 shards) — the
loop indexes past the slice length and panics with:

    runtime error: index out of range [N] with length N
    pkg/timeseries.ExtentList.spliceByTime (extent_list.go:249)

The panic surfaces as an upstream proxy-error from DeltaProxyCache
and an empty body to the caller; from a Grafana dashboard's perspective
the long-range query simply fails.

This patch:

1. Computes an upper-bound shard count from the requested range and
   maxRange before allocating, so the slice has enough capacity for
   every shard that the loop will produce (+ a small safety margin for
   the boundary tick at the end of each extent).

2. Switches from length-based pre-allocation + indexed writes to
   capacity-based pre-allocation + append. Combined with (1), large
   ranges (720 shards = 30d at 1h cadence) allocate once with no
   growth; small ranges allocate the same as before.

3. Adds regression coverage:
   - spliceByTime high shard count (6h / 1h = 7 shards) — the minimal
     case that panicked under the old fixed-4 hint.
   - spliceByTime stress 7d/1h (168 shards) and 30d/1h (720 shards) —
     real-world Grafana dashboard ranges that previously panicked.

All existing TestSplice subtests continue to pass.

Signed-off-by: thinker0 <thinker0@gmail.com>

* refactor: [timeseries] drop self-referential bug comments per PR review

Comments referenced the panic this PR fixes; once merged the historical
context becomes noise. The patched mat... (continued)

Coverage Stats

38 of 42 new or added lines in 1 file covered. (90.48%)

2 existing lines in 1 file now uncovered.

20769 of 23294 relevant lines covered (89.16%)

1.55 hits per line