tablets: store TabletInfo and TabletEntry as values

[dkropachev]

Summary

• replace *TabletInfo with TabletInfo across the tablets code path
• replace *TabletEntry with TabletEntry across the tablets code path
• update the queue, builders, metadata integration, tests, and benchmarks to use value semantics consistently
• add a GC-focused benchmark for measuring retained heap objects, retained heap bytes, and GC pause on a large long-lived CowTabletList

Verification

• go test ./...
• go test -tags unit ./tablets

GC benchmark

Compared 0c623bd (before) vs 5d43487 (after) with the included benchmark:

go test -tags unit ./tablets -run '^$' -bench '^BenchmarkCowTabletListGCShape$' -benchmem -benchtime=32x -count 5

Results:

• heap_objects: 200.6k -> 100.6k (-49.88%, significant)
• heap_bytes: 14.09M -> 12.44M (-11.74%, significant)
• gc_pause_ns: 164.0k -> 141.0k (directionally better)
• sec/op: 294.7µs -> 235.3µs (directionally better)
• gc_cycles: unchanged (6 -> 6)

The GC benchmark is included in this PR as tablets/gcshape_bench_test.go.

[dkropachev]

@sylwiaszunejko , @nikagra , please take a look

[Copilot]

Pull request overview

This PR refactors the tablets routing metadata path to use value semantics (TabletInfo/TabletEntry stored as values rather than pointers) to reduce heap allocations/GC pressure, and adds a benchmark intended to measure retained-heap/GC characteristics for a large long-lived CowTabletList.

Changes:

• Switched tablets core types and APIs from pointer-based (*TabletInfo, *TabletEntry) to value-based (TabletInfo, TabletEntry) across queueing, builders, lookups, and metadata integration.
• Updated unit tests and benchmarks to the new (value, ok) lookup patterns and value-based slices.
• Added a GC-shape benchmark for CowTabletList and updated benchmarks to the new APIs.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tablets/tablets.go	Core refactor to value semantics across tablet info/entry types, queue ops, and lookup APIs.
tablets/tablets_test.go	Updated findEntryForToken tests to new (entry, ok) return form.
tablets/cow_tablet_list_test.go	Updated COW list tests for value-based add/find/bulk-add behavior and non-pointer mutation expectations.
tablets/tablet_utils.go	Updated tablet generation helpers to return TabletInfoList values.
tablets/tablets_bench_test.go	Added GC-shape benchmark and updated existing benchmarks to value-based APIs.
metadata_scylla.go	Updated metadata describer tablet ingestion API to accept TabletInfo values.
conn.go	Updated tablet hint unmarshalling to return TabletInfo values and propagate into metadata.

Comments suppressed due to low confidence (1)

tablets/tablets.go:409

• TabletInfo.toEntry() assigns replicas: t.replicas without cloning. Now that tablets are passed around by value, callers may reasonably assume the driver takes an independent copy; mutating the original TabletInfo.replicas slice (or its elements) after calling AddTablet/BulkAddTablets could corrupt internal state and create data races. Consider either documenting the immutability requirement on the write APIs, or defensively cloning the replica slice at ingestion time (tradeoff: extra allocations).

// toEntry converts a TabletInfo to a TabletEntry.
func (t TabletInfo) toEntry() TabletEntry {
	return TabletEntry{
		replicas:   t.replicas,
		firstToken: t.firstToken,
		lastToken:  t.lastToken,
	}

[Copilot]

findEntryForToken claims to search within the half-open range [l, r), but after the binary search it can return t[l] even when l == originalR (e.g., when searching a subrange and the token is beyond all entries in that subrange). This can return an entry outside the requested bounds. Consider preserving the original r and returning (TabletEntry{}, false) when the computed index equals the original upper bound before accessing t[l].

[Copilot]

Changing TabletInfoList/TabletEntryList from slices of pointers to slices of values is a breaking change for any external code importing github.com/gocql/gocql/tablets (and for any gocql public APIs that expose these types). If this package is intended to have a stable public API, consider providing compatibility shims (e.g., new value-based types under new names and/or wrapper helpers) or clearly scoping this change to a major-version bump.

[Copilot]

FindTabletForToken now returns (TabletEntry, bool) instead of *TabletEntry, which is another breaking API change for external callers. If backwards compatibility is required, consider keeping the old signature as a deprecated wrapper (or adding a second method) to ease migration.

[Copilot]

In BenchmarkCowTabletListGCShape, token := int64((i*lookupBatchSize + j) % tabletsPerTable) performs the multiplication/modulo in int. On 32-bit platforms (or very large b.N), i*lookupBatchSize can overflow, skewing results and potentially causing negative modulo behavior. Consider doing the arithmetic in int64 (or using uint64) before the cast.

Suggested change

	token := int64((i*lookupBatchSize + j) % tabletsPerTable)
	token := (int64(i)*int64(lookupBatchSize) + int64(j)) % tokenRangeCount