Periodically prunes expired sessions from a ConduitMcp.Session.Store.

ETS-backed stores accumulate session entries indefinitely without an explicit eviction loop. Without a janitor, every successful initialize request adds a row to the session table that lives until the BEAM restarts; for a public-facing MCP server this is an unbounded memory growth surface.

Stores backed by systems with native TTL (Redis with EX, Mnesia with TTL indices, etc.) do not need this janitor.

Usage

Add the janitor to your application supervision tree:

children = [
  {ConduitMcp.Session.Janitor,
   store: ConduitMcp.Session.EtsStore,
   ttl: :timer.minutes(30),
   interval: :timer.minutes(1)},
  # ...
]

Options

• :store (required) - module implementing ConduitMcp.Session.Store that defines a cleanup/1 callback.
• :ttl - maximum session age in milliseconds (default: 30 minutes). Sessions whose created_at exceeds this age are evicted.
• :interval - interval between cleanup runs in milliseconds (default: 1 minute).
• :name - registered process name (default: module name).
• :telemetry_event - event emitted after each run (default [:conduit_mcp, :session, :cleanup]), with measurements %{removed: count} and metadata %{store: store}.
• :noun - what the swept rows are called in the debug log (default "expired sessions").

The last two exist because ConduitMcp.Application reuses this janitor for the cancellation table. Without them a consumer's [:conduit_mcp, :session, :cleanup] handler would receive cancellation evictions as session evictions once a minute, and the same count would be published under two event names - one of them wrong.