EIC code displayed by LXR master/​swf-monitor/​docs/​SNAPPER.md	Source navigation Diff markup Identifier search General search
	Version: master test Revert   Change

Warning, /swf-monitor/docs/SNAPPER.md is written in an unsupported language. File is not indexed.

 # Snapper operations in SWF
 
 Snapper is installed as a generic Django application in the SWF runtime and
 uses the existing `swfdb` PostgreSQL database. SWF owns the domain adapters,
 capture process, configuration, monitoring, and presentation.
 
 The deployment follows the same infrastructure boundary as corun-ai: Apache
 serves Django views backed by database rows, while an independently supervised
 worker performs asynchronous work. The `epicprod-ops-agent` systemd service is
 Snapper's initial worker. It invokes standalone doers and never captures state
 inside an Apache request.
 
 ## Capture scheduler
 
 The ops agent evaluates both `testbed` and `epicprod` through:
 
 ```text
 scripts/capture-system-snap.py --scope all
 ```
 
 The agent polls every 10 seconds by default. The doer reads each scope's
 effective policy from `SysConfig` on every invocation, aligns the opportunity,
 and lets the PostgreSQL capture cursor make duplicate or quiet evaluations
 cheap. The database transaction serializes each scope and locks its registered
 components in stable name order. The agent stamps each invocation before
 starting the Django doer, so process-start latency cannot move an evaluation
 into a later boundary or manufacture a coverage gap.
 
 The first read seeds these operator-visible `SysConfig` keys:
 
 | Key | Commissioning default |
 |---|---:|
 | `snapper_opportunity_seconds_testbed` | `10` |
 | `snapper_baseline_every_testbed` | `10` |
 | `snapper_capture_policy_testbed` | `testbed-v1` |
 | `snapper_opportunity_seconds_epicprod` | `10` |
 | `snapper_baseline_every_epicprod` | `10` |
 | `snapper_capture_policy_epicprod` | `epicprod-v1` |
 | `snapper_lock_timeout_ms` | `5000` |
 
 Opportunity intervals below 10 seconds are rejected. A lock timeout or another
 capture failure is recorded in the scope cursor; the next successful boundary
 creates a recovery snap. Periodic quiet and duplicate results remain in the
 bounded cursor and service journal. Material captures and failures enter the
 epicprod action stream.
 
 ## Manual capture
 
 An authorized operations publisher sends `capture_system_snap` to
 `/queue/epicprod.ops`, using the existing `prodops` namespace boundary:
 
 ```json
 {
   "msg_type": "capture_system_snap",
   "namespace": "prodops",
   "scope": "testbed",
   "created_by": "operator-name"
 }
 ```
 
 `scope` may be `testbed`, `epicprod`, or `all`. The doer targets the next
 aligned boundary when the current boundary has already been evaluated. Results
 are published as `snapper_capture_ready` on `/topic/epictopic` and recorded in
 the action stream with the requesting identity.
 
 ## Operational status
 
 System Status exposes one check per scope:
 
 - `snapper-testbed-scheduler`
 - `snapper-epicprod-scheduler`
 
 The checks read PostgreSQL cursor state only. They report stale heartbeats,
 consecutive failures, open coverage gaps, invalid SysConfig values, and the
 latest scheduler outcome. The heartbeat threshold is derived from the scope's
 opportunity interval, with a 60-second minimum.
 
 ## Web presentation
 
 The monitor mounts Snapper beneath the global System navigation. Its primary
 view state is represented by server routes rather than client-side tab state:
 
 ```text
 /snapper/<scope>/report/
 /snapper/<scope>/report/<snap-uuid>/
 /snapper/<scope>/system/
 ```
 
 `scope` is `testbed` or `epicprod`. The Report tab presents one selected full
 snap and the latest 100 history rows in a sortable table. Component envelopes
 use a generic registration-driven presentation; SWF-owned component presenters
 may add domain-specific rendering while retaining the complete component JSON.
 
 The System tab presents the effective scope policy, scheduler cursor and
 health, and the complete component registration catalog. It is read-only;
 operator configuration remains in the existing System configuration editor.
 Tab, scope, and selected-snap links are durable URL state, so reload, browser
 history, bookmarks, and external proxy links preserve the selected view.
 
 Deployments use the standard `deploy-swf-monitor.sh branch main` path. That
 freezes the generic `snapper-ai` package into the release virtual environment,
 runs migrations, updates `/opt/swf-monitor/current`, and restarts the
 `epicprod-ops-agent` worker.
 
 ## Component maintainers
 
 The full five-minute System status refresh also maintains two Snapper component
 families before the independent capture scheduler observes them:
 
 - `health` in both scopes, from the bounded System status registry; and
 - `panda` in `epicprod`, from the curated PanDA activity adapter.
 
 The PanDA adapter publishes trailing 24-hour job and task counts, current counts
 for every in-flight job state and nonterminal task state, running cores, and
 bounded target-site and task-type maps. It excludes users and individual
 records. Selected
 `refresh-system-status.py --only` runs skip the PanDA query; full manual and
 periodic refreshes publish it. A publication error causes the refresh doer to
 fail visibly while preserving the last accepted component state.

This page was automatically generated by the 2.3.7 LXR engine. The LXR team