[mgravell]

Initial IDatabase support for Redis 8.12's set-member expiration (redis/redis#15938, in 8.12-rc2) — the hash-field-TTL idea brought to sets — mirroring the HashField* family throughout. Plan document (server facts, requirement coverage, API grounding, open questions) in notes/set-member-expiry/plan.md.

Surface

Member	Command(s)	Returns
SetAddAndSetExpiry (single/array × TimeSpan?/DateTime, mirror of HashFieldSetAndSetExpiry)	SADDEX	bool — at least one member added/updated/deleted
SetMemberExpire (TimeSpan/DateTime, ExpireWhen)	SEXPIRE/SPEXPIRE/SEXPIREAT/SPEXPIREAT by precision	ExpireResult[] per member
SetMemberGetTimeToLive	SPTTL	long[] (ms; −1 none, −2 missing)
SetMemberGetExpireDateTime	SPEXPIRETIME	long[] (Unix ms)
SetMemberPersist	SPERSIST	PersistResult[]

Notable mappings, each with the sibling that grounds it:

• When.NotExists/When.Exists → MNX/MXX: all-or-nothing across the batch (one offender vetoes the whole call), which is exactly HSETEX's FNX/FXX semantics, so the same enum carries it; the XML docs spell out the batch veto.
• Bare SADDEX discards an existing member TTL; KEEPTTL keeps it — same contract as SET/HSETEX, so Expiration.CreateOrKeepTtl is reused unchanged.
• ExpireResult (2/1/0/−2) and PersistResult (1/−1/−2) match the per-member reply codes exactly, so no new enums or processors; SEXPIRE ... GT loses to a member with no expiry (no expiry = infinite), covered in docs and tests.
• Seconds-precision read twins (STTL/SEXPIRETIME) are not exposed, as the hash family does with HTTL (CommandCoverage.md re-marked not-applicable).

Keyspace notifications

8.12 fires sexpired when a member lapses, including on the subkey channels where the payload names the member (__subkeyevent@0__:sexpired, 14:presence:lobby|5:alice). No new parsing was needed: KeyNotification.GetSubKeys() on main already handles every subkey channel kind including the length-prefixed member list. The only gap was the enum entry — KeyNotificationType.SExpired = 1006 (sexpired is a side-effect event like expired, so it sits in the side-effect block, which also keeps it clear of #3261's 66/67). Unit tests cover the parse row and a GetSubKeys case on the sexpired channel.

Shared 8.12 plumbing (overlaps #3261 / #3315)

Experiments.Server_8_12 = "SER011", docs/exp/SER011.md, the NoWarn entry, RedisFeatures.v8_12_0 = 8.11.241 and the client-libs-test:8.12-rc2 Dockerfile/compose bump are textually identical to those PRs, so whichever lands first wins and the rest merge clean (the SER011.md feature bullets are the only real merge, and it's additive). All new public members are [Experimental(Experiments.Server_8_12)].

Tests

• RoundTripUnitTests/SetMemberExpiryRoundTrip — 16 exact-bytes tests (every token shape: EX/PX/EXAT/PXAT/KEEPTTL/bare, MNX/MXX, NX-style conditions, precision picking, the MEMBERS n block), all passing.
• SetMemberExpiryTests — the HashFieldTests mirror plus the set specifics (MNX batch veto, bare-form TTL discard, GT vs pinned, past-expiry delete, pinned-but-removable, WithKeyPrefix isolation): 34/34 passing (17 × RESP2/RESP3) against the rebuilt 8.12-rc2 topology (redis_version:8.11.241); they skip with the version message on older servers.
• Reply parsing reuses ExpireResultArray/PersistResultArray/Int64Array/Boolean — no new processors.
• Transaction analyzer: nothing to map — per-member TTL on a set could not be composed from any pre-8.12 MULTI shape (SADD + KeyExpire expires the key); reasoning recorded in the plan.

dotnet build Build.csproj -c Release /p:CI=true is clean apart from the pre-existing NU1903 ILCompiler advisory in toys/AotRig (fails identically on main).

[mgravell]

Reworked per review: the add-with-expiry surface is now two SetAdd overloads taking a required Expiration — SetAdd(key, value, Expiration, When, flags) and SetAdd(key, values, Expiration, When, flags), both bool, plus async — replacing the four-per-arity SetAddAndSetExpiry family (8 members → 4, no new name). Expiration encodes exactly SADDEX's expiration block (EX/PX/EXAT/PXAT/KEEPTTL); the required parameter keeps resolution away from the existing SetAdd members.

Two deliberate calls, both documented in the XML remarks:

• No SADD fallback for the no-expiry form: bare SADDEX discards an existing member's TTL (t_set.c), where SADD leaves existing members untouched — they are not equivalent, so the overload always sends SADDEX.
• Expiration.Persist is rejected with ArgumentException (SADDEX has no PERSIST token; SetMemberPersist is the removal API). Previously it would have gone to the wire as a bogus token.
• The multi-value remarks note that the server reports only whether anything changed, not a count (unlike SetAdd(values)).

The SetMemberExpire/TTL/Persist family is unchanged (GT/LT can't ride Expiration; mirrors the HashField shape). Verified locally against 8.12-rc2: all 50 SetMemberExpiry tests pass live, 525 KeyPrefixed tests pass, library builds clean.