Crate hakocluster 

hakocluster: dispatcher over N hakodb instances (Fase 2).

Full design: hakodb/hakocluster#1.

• Cluster::open[_with_config] — one Hako per data dir + full-mesh socket_sync peering (one connection per pair, i dials j > i).
• Reads (get/query) fan out round-robin across healthy instances.
• Writes (put/put_owned/delete) route to the designated writer, instances[0] by convention.
• Fase 2 adds: stagger policy (same interval + spaced opens, or per-instance intervals) and a lag guard (eject replicas trailing the writer by more than max_replica_lag_versions, re-admit at half).
• Replicated writes join each instance’s normal write/flush queue (that is SocketSync’s own behavior); the cluster adds no flush paths.

Non-unix: socket_sync compiles out (same rule as hakodb), so peering is unavailable — open with N > 1 fails closed; N = 1 works as a degenerate single-node cluster.

Modules

ffi

C ABI over JSON for non-Rust consumers (go, pascal, …).

Structs

Cluster

One clustered engine: N instances, one logical dataset.

ClusterConfig

Fase 2 cluster configuration.

Databases

N named databases in one process. Selection is by exact name; there is no default database (a silent default routes writes somewhere the operator did not choose).

DbSpec

One named database declaration for Databases::open.

Promotion

One promotion record: which epoch moved the writer where, and when. Operators use epochs to order promotions (a higher epoch supersedes); the actual fence is the read-only flags Cluster::promote sets.

ReplicaHealth

Per-instance health from Cluster::refresh_health.

Enums

StaggerPolicy

Flush-cadence stagger policy (issue #1: flush cadence is the ONLY knob — never skip the queue, no direct-flush paths).

Functions

valid_db_name

Database name gate: same discipline as collection segments ([A-Za-z0-9_-], 1–128, no __ prefix). The name becomes a sock subdir, so traversal shapes are refused here, not downstream.