Index · Glossary · Cheat sheet

The terms that Insider's documents use, with the section that gives the detail. Some words have more than one meaning. Each entry names every meaning and tells them apart.

Contents

Live observation

Session

A Live session is one observation of a running node, started with start. A node has at most one Live session, and several agents can share it. See Live observation.

A collector session (Insider.Collector.Session) is the lower-level host boundary that runs the collectors with one Recorder. A Live session runs on a collector session. See Supervised collector sessions.

Owner

The owner is the process in the target node that holds the Live session, its lease, its readers, and its marks (internal: Insider.Live.Owner). Insider starts it only where live observation is allowed. See Environment and ownership.

Observer

The observer is the short-lived, hidden node that mix insider.live or ./insider-lab live starts for one command. It calls Insider.Live in the target over Erlang distribution, prints one JSON document, and stops. See Observer node (Erlang distribution).

Component

A component is one collector that a Live session runs: calls, logger, otel, poller, or telemetry. start --component NAME chooses them. Without the option, a session runs logger, plus telemetry when the node has quiet default events. See Live observation.

Recorder

The Recorder is the bounded in-node store of a session's events. It holds 16 to 4,096 events (max_events, default 4,096) and at most 4 MiB. When it is full, it evicts the oldest event, and reads report the loss as a gap. See Cursors and gaps.

Receipt

The word has three meanings:

  • Event receipt. The number that the Recorder gives each event it admits, consecutively from 1. Cursors and marks are receipts. See Cursors and gaps.
  • Cleanup receipt. The receipt (insider.collector-session-receipt/v1) that a stopped or expired session reports, as proof of its cleanup. See Stop and check the cleanup receipt.
  • Sentinel receipts. The bounded records that Sentinel keeps of a confirmation, a verification, or a deduplication ledger entry. See Collector-to-Sentinel handoff.

Cursor

A cursor is the receipt that a read starts after. Cursor 0 is the start of the session. Each read returns next_cursor, the cursor for the next read. Cursor latest is the newest receipt, so its batch is empty and its next_cursor starts a tail. See Cursors and gaps.

Gap

A gap is a known loss in a read: a %{"reason", "count", "scope"} entry in "gaps". An empty list means no known loss. Insider never loses events silently. See Cursors and gaps.

Reader

A reader is a named agent that reads a session (--reader NAME). The session tracks each reader's cursor and view, and status lists them. A stop while another reader is active returns readers_active, unless the caller adds --force. See Several agents on one session.

View

A view narrows what a read shows, by source, kind, and minimum severity. It never changes what the session records. A per-read view applies to one read (--view-source, --view-kind, --view-min-severity). A stored view is a named reader's own view, which later reads by that reader use. See Filters, summaries, and views and Several agents on one session.

Mark

A mark is a labelled point in a session, made with mark --label just before an action. It records a live.mark event at the newest receipt. Reads with --since-mark start after it. A session keeps its newest 256 marks. See Pairing marks.

Window

The word has three meanings:

  • Digest window. The receipts that a digest, baseline, or trends call reads, set by --since or --since-mark. See Digest.
  • Mark window. Everything after a mark, up to the newest receipt. See Pairing marks.
  • Retention window. The events that the Recorder still holds, plus the traces that a session pins outside the Recorder. See Pin a trace.

Baseline

The word has three meanings:

  • Known-normal baseline. The anomalies of a window that baseline marks as normal for a session. Later digests list them as known and lead with new ones. See Known-normal baseline.
  • Load-test baseline. The uninstrumented phases of a synthetic load run, which the instrumented phases are compared with. See Synthetic load and soak evidence.
  • Oracle boundary. The point in test/test_helper.exs, after any intended suite-baseline configuration, where an agent installs an oracle. See Value-free application-env oracle.

Lease

The word has two meanings:

  • Live session lease. The time a session lives without a renewing call: lease_ms, 1,000 to 900,000, default 120,000. When it runs out, the owner stops the session as expired. See Lease.
  • Sentinel sandbox lease. The owner lease on the resources of one Sentinel run. It is complete only when an independent owner inventory is empty. Otherwise it stays quarantined. See Collector-to-Sentinel handoff.

Disclosure profile

The disclosure profile decides what a Live session lets leave the node: rich (payloads, rendered and bounded) or strict (opaque, shape-checked values only). The host chooses it, and a caller can only narrow it. Do not confuse it with a load profile or an oracle host profile. See Disclosure profiles.

Masking

Masking replaces credential-shaped values with [REDACTED:<reason>] in every profile. It works by key name and value shape. Each mask adds a mask record to the event's "redactions", never the value. Redaction of portable evidence is a separate, policy-driven step. See Credential masking and Redaction boundary.

Exemplar

An exemplar is a full event that a summary keeps as an example: up to three of an interval's slowest events and its first error. The summary names each by receipt, so an agent can open it. See Filters, summaries, and views.

Lab and Sentinel

Lab

The Lab is tools/insider_lab, a separate Mix project, published as insider_lab, for development and disposable environments. It holds the observer task (mix insider.live), its escript, the candidate inventory, the oracles, and Sentinel. A production release cannot load Lab modules. See Repository layout.

Sentinel

Sentinel is the Lab-only part that takes a capsule and tries to prove its failure on a clean target. It confirms, verifies, makes a regression, minimizes, and bisects. Then it hands off to a coding agent for a draft proposal. See Proof workflow.

Capsule

An incident capsule (Insider.Incident.Capsule) is a deterministic, redacted artifact of one incident. It holds sanitized events, a finding, provenance, and digest-only artifact references. It accepts at most 64 events, 32 artifact references, and 1 MiB of canonical JSON. See Incident capsules.

Candidate

A candidate is one entry of the inventory that mix insider.candidates writes: a test with static isolation or nondeterminism signals. It is a heuristic, never a confirmed defect. See Agent-requested ExUnit candidates.

Oracle

An oracle is a fixed Lab-only check that answers one question about a selected test run. It reports without values. Examples: whether a test changed an application-env key, left a Telemetry handler behind, or changed OpenTelemetry global state. See Oracle runner.

Production readiness and load

Proof of concept

A proof of concept is an early, deliberately small implementation that shows an approach works. Insider as a whole is one, so its protocol, APIs, and package layout are not stable. See Production boundary.

Load profile

A load profile is the fixed-seed synthetic workload that tools/insider_load runs without and with a collector session. It is not a disclosure profile. The threat model's "future production profile" is a third meaning: the defaults a production deployment would need. See Synthetic load and soak evidence.

Production readiness

Production readiness is the evidence that production observation would be safe. It grants no authority: staging and production observation stay denied until separate human decisions enable them. See the threat model, the production-readiness dossier, and Production boundary.