caddymodule

hazedb as a Caddy module

This package wraps hazedb as a Caddy HTTP handler, so a single Caddy binary serves your site and an in-process SQL store — no separate database process, no network hop.

It exposes five endpoints. POST /query and POST /exec take a JSON body {"sql": "...", "args": [...]}; the GET routes take URL parameters.

GET /meta takes no parameters; it reports the table count, the configured max_bytes cap (0 = unlimited), the store-wide total_rows / total_approx_bytes / total_tombstones, and per table the row / column / index counts, an approximate in-RAM byte size, and tombstones — for dashboards and health checks. The byte sizes are estimates (cell payloads plus modeled overhead, biased slightly high), not exact accounting.

Tombstones are rows deleted but whose memory slot is not yet reclaimed. A background sweeper compacts shards that have gone mostly dead, so tombstones is a gauge — it rises with deletes and falls as the sweeper runs, rather than only resetting on restart. A momentarily high tombstones / (rows + tombstones) fraction between sweeps is normal; a persistently high one means deletes are outrunning the sweeper. (Scan cost is independent — partitioned scans stay proportional to live rows regardless.)

Byte cap. Set max_bytes (below) to bound the store's RAM. An INSERT that would push total_approx_bytes past the cap is rejected with HTTP 507 (Insufficient Storage); the store never auto-evicts, so the client frees space with DELETE / DROP TABLE. total_approx_bytes vs max_bytes from /meta is the headroom gauge.

args is an optional positional list for ? placeholders. JSON → SQL value mapping: number → INT, bool → BOOL, null → NULL, string → STRING, except a string in canonical UUID form (8-4-4-4-12 hex) → UUID — so you can address and insert UUID columns from JSON.

Build a Caddy binary with the module

Uses xcaddy. Until the core hazedb version that ships db_registry.go + wire.go is tagged, build against the local checkout (the submodule's replace points at ../):

# from a checkout of this repo
xcaddy build \
    --with github.com/VeloxCoding/hazedb/caddymodule=./caddymodule \
    --with github.com/VeloxCoding/hazedb=.

After the core is published, the plain form works:

xcaddy build --with github.com/VeloxCoding/hazedb/caddymodule

Configure (Caddyfile)

:8080 {
    handle /db/* {
        hazedb {
            wal            on                        # on (default) = durable (crash loses ≤~0.5s); off = memory-only
            wal_path       /var/lib/hazedb/wal       # optional: override the WAL directory. Default <caddy-data-dir>/hazedb/wal
            companion_path /var/lib/hazedb/hazedb.db # optional: on-disk SQLite companion (events/health always, + data mirror when wal is on). Default <caddy-data-dir>/hazedb/hazedb.db (or next to wal_path)
            init_sql       /etc/hazedb/schema.sql   # CREATE TABLE + seed, run once at startup
            registry_name  default                  # name the *DB is published under for the PHP extension
            max_body_bytes 4194304                  # POST body cap for /query and /exec (default 4 MiB)
            max_bytes      1073741824               # cap the store's RAM (1 GiB); over-cap INSERT → HTTP 507. 0/unset = unlimited
        }
    }
}

All subdirectives are optional. By default the store is durable (wal on): with nothing set, the WAL and the SQLite companion both live under <caddy-data-dir>/hazedb/ (e.g. /var/lib/caddy/hazedb/ under systemd). Set wal off for a memory-only store. The WAL has one durability story — writes seal to disk within ~0.5s, so a crash loses at most that window; there are no durability levels or fsync modes. Tables are created at runtime: put your CREATE TABLE statements in the init_sql file, or POST them to /exec.

Schema / tables

hazedb creates tables at runtime, so the module opens with an empty schema. Define tables one of two ways:

• init_sql — a file of ;-separated statements run once at startup (typical: CREATE TABLE ... plus any seed rows). It runs as a trusted script, so a ; inside a string literal is safe and seed rows may use inline literal values.
• POST /exec — send CREATE TABLE ... like any other write.

Sharing the instance with PHP

The module publishes its *DB in hazedb's process-wide registry under registry_name (default "default"). The FrankenPHP/PHP extension (addons/frankenphp-ext) looks up that same name, so PHP calls and HTTP calls hit one identical database — no second copy.

WAL + config reload caveat

With wal_path set, a Caddy graceful config reload runs the new handler's Provision (which opens the WAL file) before the old handler's Cleanup (which closes it) — briefly two writers on one file. Memory mode reloads cleanly. For durable deployments, prefer a full restart over a graceful reload when changing this handler.