Read-After-Write Consistency

The Guarantee, Precisely Defined

 READ-AFTER-WRITE (session consistency):
   a user ALWAYS sees their own completed writes,
   regardless of which replica serves their reads.

 scope matters:
 - PER USER/SESSION:  Sarah sees Sarah's edit immediately
                      others may lag (that's fine!)
 NOT global linearizability — that costs far more
 and users can't tell the difference.

Why It Breaks by Default

 write → primary ──async──► replicas (800ms behind)
 read  → LB picks any replica → STALE for that user

 broken flows in every naive deployment:
 - rename profile → refresh → old name
 - post comment → it's not in the list
 - upload avatar → old face stares back

Implementation 1: Sticky Window

 after a WRITE, mark the session:

 redis.set("raw:u912", now(), ex=10)      # 10s window

 on READ routing:
   if session_wrote_recently(user):
       route to PRIMARY
   else:
       route to any replica

 properties: trivial to build; slightly over-routes
 (primary load rises with write-active users); window
 must exceed worst-case replication lag.
 THE pragmatic default.

Implementation 2: Version/LSN Tokens

 writes return a POSITION token:

 POST /profile → 200 { ..., "_rv": "wal:8f3a92" }
 client echoes token on subsequent reads:
 GET /profile?_rv=wal:8f3a92

 router logic:
   if replica.replay_lsn >= _rv:  serve from replica ✓
   else: wait briefly / fall to primary ✓

 precise: primary touched ONLY until replica catches up,
 then traffic returns. more machinery; best at scale.
 managed equivalents exist (DynamoDB consistent-read tokens).

Implementation 3: Affinity Routing

 route each USER consistently:

 hash(user_id) → replica group
 their writes' replicas are likely caught up with them;
 plus sticky-window for safety right after writes.

 bonus: per-user cache affinity comes free
 (same replica = warm cache for that user's slice).

Choosing Per Endpoint Class

Flow	Mechanism
Profile/settings pages	Sticky window
Order confirmation	LSN token (correctness matters)
Feeds/browsing	Nothing needed (staleness fine)
Auth-critical checks	Primary always

 resist applying mechanisms globally:
 blanket primary-routing after ANY write recreates
 single-node read scaling. scope tightly, measure, relax.

Interview Framing

“User complains their edits don’t appear” scored diagnosis: replication-lag + replica-routing as root cause, fix menu presented with tradeoffs (sticky=cheap, tokens=precise), scoped-per-endpoint-class table. The precision of the guarantee definition (“session-scoped, not global”) is itself a seniority marker — say it explicitly.