perf(tsdb/index): allocation-free postings offset table scan

[simonswine]

Summary

newReader spent ~96% of its allocations inside ReadOffsetTable: every entry materialised a []string plus two string clones for name and value, even for the ~31/32 entries that symbolFactor sampling immediately discards.

This PR introduces readPostingsOffsetTable, a sibling of ReadOffsetTable that yields (name, value []byte) slices aliasing the mmap buffer instead of allocating strings. The newReader callback:

• uses yoloString for map lookups (zero allocation),
• defers string(name) / string(value) conversion until an entry is actually retained, and
• tracks the "last value" candidate as raw byte-slice aliases, cloning only at flush time.

Impact

This path represents 66% of the allocation counts in query-backends (and a very small amount of compaction workers)

Benchmark

BenchmarkNewReader with 10 label names × 200 values each:

             │      sec/op       │      sec/op      vs base  │
NewReader-12      305.1µ ± 1%       124.2µ ± 1%    -59.30%

             │       B/op        │       B/op       vs base  │
NewReader-12      270.7Ki ± 0%      169.8Ki ± 0%   -37.30%

             │     allocs/op     │   allocs/op    vs base    │
NewReader-12       6075.0 ± 0%       244.0 ± 0%   -95.98%

Allocation count drops from 6 075 → 244 (−96%), wall time from 305 µs → 124 µs (−59%).

Notes

• readPostingsOffsetTable is unexported and used only by newReader and PostingsRanges.
• This builds on refactor(tsdb/index): drop FormatV1 reader support #5158 (FormatV1 reader support removed), which clears away the V1 branch that previously shared ReadOffsetTable.
• Inspired by the newer version of Prometheus code base: https://github.com/prometheus/prometheus/blob/43e5fc6a248c1a73d30dff86428f74d11aedbedc/tsdb/index/index.go#L1297-L1302

Test plan

• go test ./pkg/phlaredb/tsdb/index/... -run . -bench BenchmarkNewReader passes and shows regression-free numbers
• Existing index reader tests (TestNewReader, table-driven cases) pass unchanged

---

Note

Medium Risk
Touches TSDB index reading/parsing and introduces more buffer-aliasing ([]byte/yoloString) to avoid allocations, which could cause subtle lifetime/aliasing bugs if misused despite being localized and covered by tests/bench.

Overview
Reduces NewReader startup allocations/time by replacing the generic ReadOffsetTable postings scan with an allocation-free readPostingsOffsetTable that returns (name, value []byte) slices aliasing the index buffer, and only materializes string values for sampled/required entries (including correct handling of the per-label “last value” entry).

Updates PostingsRanges to use the new postings-table iterator, adjusts the label-indices test to decode the table directly, and adds BenchmarkNewReader to track open-index performance/regressions.

^{Reviewed by Cursor Bugbot for commit e07c1e0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[aleks-p]

Great improvement!

[aleks-p]

nit, maybe we can remove the mention of mmap here to avoid confusion, it looks like the data is always heap allocated (and it shouldn't matter in general)

[simonswine]

Likely coming from the prometheus context 👍

[simonswine]

Fixed in e07c1e0