perf(db): pre-size memtable slice in getMemTables

[shaunpatterson]

Summary

The memtable slice returned by getMemTables was previously grown via append-from-nil, triggering a growslice on every call (and additional growslices for each cap doubling boundary crossed). The final size is known up front: 1 mutable (unless ReadOnly) + len(db.imm) immutables. Pre-sizing collapses all growslice events into a single make().

Motivation

Profiled in BenchmarkRollupKeyIterator: the original append-from-nil allocation (~1.3 allocs/op from growslice) is eliminated; in its place is a single make() matching the final size.

In real dgraph workloads with 6 memtables (1 mutable + 5 immutable), the original code grew the slice through 4 cap boundaries (0→1→2→4→8), so this saves 4 allocs per iterator construction. Each rollup cache miss constructs one iterator.

Behavior

Preserved. The only observable difference is that an empty result returns a non-nil zero-length slice rather than nil. All three callers (db.get, txn.NewIterator, stream_writer) use len() or range, neither of which distinguishes the two.

Test plan

• go test ./... passes
• Existing memtable/iterator tests cover the path

🤖 Generated with Claude Code