PlanetScale Neki: cooling hot shards when tenant_id sharding breaks

PlanetScale published an engineering deep-dive (2026-09-28, Simeon Griggs) on handling hot shards in Neki (sharded Postgres): when even tenant_id distribution fails under whale tenants and new cross-tenant product needs, teams can scale a shard, isolate a whale into its own key range, or reshard selected tables by a better key—Reshard copies + streams while serving, then switches traffic without taking the app offline (blog).

This is a Desk Bot devops/postgres briefing. Prefer short PlanetScale-blog paraphrase—no dashboard/UI dump. HARD: do not invent a Neki GA date beyond the post’s “since the launch of Neki” framing.

The problem

Sharding on tenant_id starts even, then hot tenants (size/volume) and cross-tenant product needs make “even tenants” worse than “even data.” Neki’s router uses a declarative data topology for write placement and read shard lookup (blog).

Three mitigations (as stated)

1. Scale up — Each shard is its own Postgres cluster; assign a unique config profile and vertically scale the hot shard to buy time.
2. Isolate whales — Tighten key_ranges so a hot tenant_id hash lands on a dedicated shard (example: narrow start/end around the whale). Requires reshard onto new shards (cannot reuse source shards while copying). Example key_ranges JSON in the post is illustrative, not a production recipe.
3. Reshard some tables — Change shard key per table from access patterns (e.g. messages by channel, not workspace); keep unsharded tables that need whole-scope queries (e.g. users).

Online path: After topology change, Reshard copies existing rows, streams ongoing writes while the source serves, then switches reads/writes without app downtime (blog).

Soft locks

• Declarative vs automatic: Explicit topology (which tables, which key, which ranges) vs opaque auto-placement—post pitches this as agent/human-readable; attribute PlanetScale.
• “Fastest cloud Postgres,” agent-friendly topology — vendor narrative; soft-attribute only.
• Slack Vitess workspace→channel pattern is a third-party historical analogy—do not imply Slack runs on Neki (blog).
• Out of scope: full Neki feature dump, Vitess MySQL vs Postgres deep dive, pricing.

Who should care

Multi-tenant AI SaaS teams hitting whale-tenant skew on sharded Postgres should start at the PlanetScale hot-shards post—keep the three mitigations + online Reshard as stated, treat Slack as analogy only, and leave Neki GA dating to “since the launch of Neki.”