An application server runs dozens of background executors: samplers, watchdogs, cache evictors, change-data-capture readers, the scheduler, the AI agent tools. Until this release an operator saw them as thread names chosen by whoever wrote the code. Now each one is named, says what it is for, is built from what it promises, and is held to it.
Named in three levels
- Component The subsystem, from a fixed list: core, jvm, jdbc, cdc, cache, session, security, cron, mcp, ai, backup and others.
- Executor What it is for, such as
jvm.rate-samplerorjdbc.node-watchdog; its threads carry the same name. - Owner The code that made it, kept as a column of its own so grouping no longer depends on class names.
- Last run Every scheduled job records its schedule in words, when it last started, how long it took and its last error.
The policy is the input
- Declared once Mission, period, longest run, lateness, queue limit, what happens to a task it cannot take, what a failed job does, and what it feeds.
- Built from it The registry sizes the pool, its queue and its rejection rule from the declaration; an executor without one is refused, and a build-time test fails on any that is undeclared.
- Stopped by it At shutdown each executor drains within a time budget, is interrupted or is abandoned as declared, in an order that stops entry points first and the log queue last, before the connection pools close.
- Checked against it A live table sets each promise beside what the executor does now, one row per aspect with a verdict of OK, Watch, Breach or Undeclared; each change of verdict is kept as an event.
- Virtual threads observed Pinning of virtual threads over 20 ms is attributed to its executor and frame through Flight Recorder.
The change reached the core, the web tier, the AI agents and the startup module in the same week. An operator reading the executor list now reads a mission statement and a verdict, not a thread dump.