One allocation, two reads.

The old table was chained: a bucket array of pointers, each entry a separately allocated dictEntry of three pointers, the key its own allocation — four memory reads to resolve one key, plus two per collision hop, plus a second table during rehash.

Its cost ledger, per key: a bucket-pointer share, a 24-byte dictEntry plus allocation overhead, a separate key allocation, and — with a TTL — a second entry in the expires dict. Every line of that ledger is a pointer to chase, and every pointer is a cache line the CPU may not have.

The 8.1 design inverts the layout around the cache line: buckets of 64 bytes holding up to seven entries, an 8-byte metadata section — one child-pointer bit, seven presence bits, and seven one-byte secondary hashes built from the bits not spent addressing the bucket, so mismatched slots are skipped without touching their keys at a 1-in-256 false-positive rate. The dictEntry is gone; key and value embed in the serverObject, one allocation per entry, two memory reads per lookup.

Incremental rehash, the SCAN guarantee, and random sampling — the features that ruled out adopting Swiss tables wholesale — are kept by construction. And one source detail matters to ID design: every key embeds regardless of length, and keys of 128 bytes or more pre-reserve an expire slot (KEY_SIZE_TO_INCLUDE_EXPIRE_THRESHOLD, object.c:47 in the 8.1.8 tree this chapter built).

The new table charges a share of 64 bytes across up to seven entries and one object that already contains key, value, and, when set, an inline 8-byte expire. The publication's headline is roughly 20 bytes saved per pair, roughly 30 with a TTL — and the claims reproduce on this project's own keys: 23 to 39 bytes saved, 29 on the TTL row.

fmt keylen redis7 valkey81 saved
brd14 14 88 65 23
uuid16 16 104 65 39
brd14+ttl 14 128 99 29

jemalloc allocates in 16-byte classes in this range, so key length stops mattering within a class and matters sharply at its edges: an identifier's storage cost is a staircase, and the design question is which step each format stands on. The five measured key lengths plot the steps:

The staircase is the durable model: the 9.x line extends the same bucket structure with SIMD lookups and batch prefetching, so class-edge thinking outlives this release.