Batch gaps and unsealed days

What you're seeing

One of these, on the Home page, the Brief or the Findings page:

  • Batches never arrived: a Medium finding, one per source, such as "7 batches never arrived in 6 gaps; the latest was detected …. Evidence inside a gap is missing, not empty." On Home it also appears in the feed as "Batches #22211–#22213 never arrived (Cruz)", and the Brief counts batch sequence gaps in the window.
  • Day not sealing / unsealed days piling up: the Home page shows recent days without a root, and the number of sealable days grows.

What it means

Two different problems that look similar:

  • A gap means an emitter's batch numbers jumped: batches N and N+3 arrived but N+1 and N+2 did not. The chain is intact before and after the hole; the hole is the evidence that something was dropped. Records inside a missing batch are missing, not empty, so any conclusion about that source for that time has a hole in it.
  • Unsealed days mean ingest is unhealthy in a sustained way: records are arriving late, being rejected, or not arriving. Sealing waits for the window to close; days pile up only when that keeps failing.

A late batch is accepted. When every batch in a gap's range has arrived, the gap closes and the finding resolves itself. The record of the skip stays as history.

Reading a gap in the product

On the Findings page, click the Batches never arrived card. A window opens with the days that have gaps as filters, and a table of the batches below. Pick a day, then click a batch to open its proof thread: the batch received before the gap, the missing range, the batch received after, and whether that day has been sealed, with a line from the batch up to the finding. It answers what is missing, whose, where, so what, and now what for that one batch.

Whose names the emitter or collector by the name you gave it, with the end of its id so it can be matched, and the agents that source delivers for, most records first. A batch that never arrived has no records to read an agent from, so the agents named are the ones that source has delivered for before: where the missing batch would have gone, not a proven list. The agents' activity is what may be incomplete. Records on that source that name no agent are counted separately. If the source has no name, it is shown as its kind and the end of its id, and naming it makes every gap on it easier to read.

To look at one source's gaps only, use the Source filter in the header. See Sources, filters and comparing.

Why batches go missing

  1. The emitter restarted and lost its queue. An ungraceful shutdown, a spool directory wiped by a deploy, or disk pressure deleting queued batches. Usually one long range.
  2. Delivery failed at the moment a batch was sent, and the emitter did not retry it. The network dropped, or the ingest endpoint was briefly unavailable. Provantix is restarted for releases and the endpoint is unavailable for a few seconds, occasionally up to a minute; the changelog lists releases. An emitter configured with a durable spool keeps a batch until the platform acknowledges it and resends it in order, so nothing is lost and no gap appears. An emitter running without one can lose a batch it fails to deliver. This usually looks like many small gaps, often one batch each, clustered at the same minutes.
  3. The emitter takes a batch number before it sends, and a failed send uses the number up. If a retry then takes a fresh number, one failed delivery leaves one to four numbers that never arrive, even when a later retry succeeds. An emitter that sends every event as its own batch burns numbers often. A spool directory does not prevent this, because it protects queued events and not numbers already used. The fix is that a retry resends the same signed batch under the same number until the platform acknowledges it, which is how the Provantix SDK works. This usually looks like many short gaps of one to four batches, with longer runs in bursts during an outage.
  4. A batch was rejected. Check Connectors for rejection codes on the source. A sequence conflict (two processes shipping under one emitter id) rejects batches and looks like a gap.

What to do for a gap

  1. Open the finding and confirm batches after the gap are landing. If so, the gap is historical, not ongoing. If the latest gap was detected minutes ago, it is still happening.
  2. Match the pattern to a cause. One long range points to cause 1. Many single-batch gaps at the same minutes point to cause 2: compare their detection times with the platform's release times, and with your own network or host events. Many gaps of one to four batches, especially of exactly the emitter's retry limit, point to cause 3: look at how the emitter assigns a number to a batch that fails and is retried. The AI Insider counts short and long gaps for you.
  3. Fix the cause at the emitter. For cause 3, make a retry resend the same signed batch under the same number. For the others, configure a durable spoolDir on a persistent volume so an unacknowledged batch survives a restart and a failed send. See Node.js SDK and Troubleshooting.
  4. Do not rewind or reset the emitter's sequence. A continuous chain with one declared gap is stronger evidence than a silently restarted one.
  5. If you can recover the missing batches from the emitter's spool, resend them. They are accepted late and the gap closes.
  6. Resolve the finding with the cause, for example "emitter had no spool; configured spoolDir on a persistent volume; batches continuous since 14:20". The gap stays visible in the chain history; that is correct, it is part of the record.

What to do for unsealed days

Treat it as an incident, not a finding to dismiss:

  1. Open Connectors and check each source's last delivery and rejection codes. A source with last delivery never or a wall of rejections is the likely root.
  2. If a source is silent, follow Emitter silent.
  3. Once deliveries resume, sealing catches up on its own: the sealer runs on schedule and roots each closed day. "Next seal in 6h" on a current day is normal; six unsealed past days is not.
  4. A day is sealed over the batches that arrived. A gap does not stop a day sealing, but the root then covers what arrived and the proof thread says so.

What good looks like

Coverage shows every closed day rooted, the integrity strip shows no open gaps, and any historical gap has a resolved finding naming its cause. A source that delivers reliably through a platform release shows no new gaps afterwards.

See also: Your first Brief says attention for where gaps sit in a first reading of the Brief.