Reduce the worst-case concurrent ClickHouse memory of the internal
platform-analytics route (all 17 CH queries fire in one Promise.all on
the shared admin user, against a 9 GB per-user cap).
- Use sipHash64(user_id) as the distinct key in the uniqExact/uniqExactIf
aggregates (dauSeries, sparkByProject, mauProjects, activeByProject).
Exact at this scale (64-bit, negligible collision prob over 1M users),
verified byte-identical; -40% to -61% peak memory per query.
- Sample the new/retained/reactivated activity split at 1-in-4 users
(consistent cityHash bucket on both subqueries, counts scaled x4). The
split's window function + all-history scan made it the heaviest query
(~1.3 GiB at 1M users / 50M events); sampling cuts it ~78% for a ~0.4%
mean error. Dashboard chart now notes it is a sampled estimate.
Adds the benchmark/optimization harnesses used to validate these changes
(seed isolated bench_pa DBs, measure peak memory, verify result equality).
