RPC and shard counts per operation, plus the two staleness cases
Count shards, not milliseconds
4 base shards, 2 index shards; every count below comes from the Java run
operation
local index
global index
write, 2 terms
1 shard
1 base + 2 index
read, 1 term
4 shards, scatter
1 index + fetch
read, 2 terms
4 shards, scatter
2 index + fetch
partial index write
same-shard txn
AND misses row
stale index entry
revalidate ids
delete wins
counts are RPCs, never latency: local timing cannot predict networked p99 — say so in review
the reader revalidates every id against its base shard; tripwire on fan-out equal to shard count
legend: local pays on reads, global pays on writes — pick the side your workload forgives