Collector JSONL records

Collector JSONL holds the process samples the Foxborne collector takes on a vehicle's companion computer. Each line is one JSON object: one sample of one process. Files arrive with the collector's uploads, or through Import files like the other formats.

At the default rate of 1 per second, each sample records CPU, resident memory and restart count for one watched process. The memory_pressure rule reads these records.

Process sample fields#

Record for a missing process#

When a watched process is not running at a sample, the collector still writes a record. It sets present to false and pid to null, and it carries no memory, CPU or restart values.

{
  "t_utc": "2026-09-21T03:14:08.402Z",
  "boot_id": "5d0c88e1",
  "proc": "perception_node",
  "pid": null,
  "present": false
}

The console titles this event perception_node sample missing, process gone. In the example dataset, samples of a running process carry no present field.

In INC-0139, chrony never synchronized, so t_utc reads the companion's own clock: 03:14, during a flight that ran from 14:05 to 14:23 UTC. The manual alignment proposed for that incident puts the offset at +11:03:04.916, bound ±1.04 s.

Example record#

This is the first of the three samples behind the memory_pressure finding in INC-0142, row 2,811 of uas04-orin-samples.jsonl:

{
  "t_utc": "2026-09-24T14:32:01.204Z",
  "mono_ns": 1149990000000,
  "boot_id": "9b2e4c1d",
  "proc": "perception_node",
  "pid": 2213,
  "rss_kib": 6336512,
  "cpu_pct": 187.4,
  "restarts": 0
}

• rss_kib 6,336,512 is 6,188 MiB, at or above the rule's 6,144 MiB threshold. The next two rows read 6,402 and 6,655 MiB, and the rule confirmed the finding at 14:32:03.204.
• mono_ns sits on the journal's clock. The kernel's journal line reporting the kill of pid 2213 carries __MONOTONIC_TIMESTAMP 1152904000. That is 2.914 s after this sample on the same clock, and the two wall times differ by the same 2.914 s.
• After the restart, samples carry pid 3398 and restarts 1, matching the journal's "restart counter is at 1".

The evidence inspector shows the same record, indented, under Original record, with Rows 2,811 to 2,813 under Location.

What the collector records on every sample#

What the collector reads, in the Collector policy drawer under Sources, Vehicle collectors, lists the clock data the collector keeps. Its Clock row reads: "Wall time, monotonic time, boot ID and chrony status on every sample. SYSTEM_TIME pairs from 0.8."

The Clock column of Vehicle collectors shows each vehicle's chrony state, such as Synced +0.8 ms or Not synchronised. Collectors on 0.7.3, the older version in the example dataset, show No SYSTEM_TIME beside the version. The collector never records environment variables, process arguments or camera frames.

How Foxborne recognizes the format#

Import files reads the first 4,096 bytes of each file and runs these checks in order. The first match sets the format.

A file that matches none is rejected with Not a format Foxborne reads. Expected PX4 ULog, journalctl JSON, collector JSONL or receiver JSONL.

The file name counts in one case only: a file named .ulg without the magic bytes is quarantined rather than rejected. Because the receiver check runs first, a file with "recv_utc" in its first 4,096 bytes imports as receiver JSONL.

The clock Foxborne assigns#

Source files on the run page lists process samples with these values, here from run R-0931 in the example dataset:

The journal and process samples read Same clock under Ordering bounds between sources, so their events are always in order against each other.

When chrony never synchronized on that boot, the journal has no anchor and the samples have none either. In run R-0922 both show None: chrony never synchronised, and the timeline keeps them on the companion's own clock. Collector 0.8.1 records SYSTEM_TIME pairs, which give companion evidence a shared anchor with the flight log.

Where the console shows these records#

• Timeline: the Process samples lane draws each process as a series in MiB, with the rule's threshold line, such as Threshold 6,144 MiB.
• Evidence inspector: Original record holds the record, and Location gives its rows.
• Event page: Memory around this event charts the samples from 8 s before to 4 s after the event. Breadcrumbs collapse samples per PID, such as Process samples: perception_node, pid 2213.

Related#

• The companion collector and Set the collector policy cover what the collector samples and how often.
• memory_pressure is the rule that reads these records.
• Clocks, anchors and error bounds explains the shared boot ID anchor.
• Receiver JSONL records documents the ground receive log.