iopsystems/rezolus v5.22.0

on GitHub

Added

• Reader: a group table whose slots the identity index describes is read
  through it. A group table's columns are its slots, and a slot changes hands
  — a task exits and another lands in its BPF map slot, a cgroup id is
  recycled — so a column's own labels can only say what the slot meant when
  the segment was written. The reader now replays the stream's caller_rows
  into per-slot occupancy spans and cuts each column into one series per
  occupant, labelled with what the index says, with the occupant's __uid__
  on it. On archives that carry identity in both places today the two read
  paths agree series for series; once the writer stops copying identity into
  column metadata, the index is where the labels live. (#1280, #1284)
• Recorder: record --stream restates every stream's slot set into the
  archive every 300 s (the seal policy's age bound), so a rolling buffer can
  evict identity history and a reader of a long recording replays one
  restatement period rather than everything since connect. Retention cuts a
  stream's caller_rows back to the latest full restatement at or before the
  row cutoff and never past it; a stream with none on record keeps its
  history. (#1281)

Changed

• Reader: a .rez table's segments are fetched from the archive as a query
  touches them, through a cache bounded by --cache-size-mb, instead of
  every segment's bytes and parsed footer being held for the reader's life.
  Open reads each footer once for the identity indexes and a per-segment time
  span, then lets it go. On a 1.3 GB, 9.6-hour archive whose per-task table
  had 159 segments of up to 2,851 columns, the viewer's resident size after
  building its dashboard went from 4.1 GB to 0.25 GB and mcp describe-metrics from 4.4 GB to 0.25 GB. Needs metriken-query 0.27.0.
  (#1283)
• Reader: rate() and irate() read each series as a sample stream and keep
  one interval's worth of it, so a query over many series costs the series it
  is looking at rather than the whole table. On the same archive,
  sum(rate(task_cpu_usage[1m])) over 6,644 task series peaked at 11.9 GB
  before and 0.98 GB after, with identical output. Needs metriken-query
  0.29.0. (#1286)
• Reader: composing a recording (systemslab's path) hands out one lazy child
  per table, built from the catalog the reader probes at open; a table opens
  the first time a composed query names one of its metrics, and composed
  rate() queries stream. Composing this archive used to open every table's
  every segment footer before any query. Needs metriken-query 0.30.0.
  (#1287)

Fixed

• Reader: replaying a stream's identity index diffed the whole live slot set
  after every entry and serialized every slot for a state hash the reader
  never reads. On a ten-minute recording under task churn (248,511 entries)
  a query took 20 s; it takes 0.6 s. (#1282)