Skip to content

Releases

boxd 0.4.75

Choose a tag to compare

@github-actions github-actions released this 22 Nov 15:26
  • box ps without --pod, when the build server has no such box, now adds one line: the name it asked is this worktree's own
    box, and a box started with --pod <name> is asked with box ps --pod <name>. A caller whose boxes all have names read the
    bare "no box" as "my box is gone".
  • Skill text: a tool kept in /cache may still need its system libraries in each new box.

boxd 0.4.74

Choose a tag to compare

@github-actions github-actions released this 22 Nov 12:16

Two faults found within the hour of 0.4.72 and 0.4.73, on real hosts.

  • pod log for a running run started without --alive printed the shell's own complaint on standard error at every call (boxd-pod: line 949: …/runs/<run>.alive: No such file or directory), so box log <run> --state and the new --quiet-limit showed it each time they asked. The answer itself was right. It is silent now.
  • The host watcher's new zram finding compared its two numbers as text, so a peak of 24 MB counted as more than a 3934 MB swap limit and a false finding was written. They are compared as numbers.

boxd 0.4.72

Choose a tag to compare

@github-actions github-actions released this 22 Nov 11:38

The memory that compressed swap really takes is in the host report, and the host watcher says when it passed what the cap leaves for it.

  • pod report: two more words end the host line, zram_used_mb and zram_peak_mb: the memory the zram devices hold now and at their peak since boot (mem_used_total and mem_used_max of /sys/block/zram*/mm_stat, summed, megabytes rounded down), a dash for each without a zram device. They count the whole host, not only the build group. group_swap_used_mb stays the logical size of what the group swapped out; the new words are what that costs in memory.
  • The memory cap leaves half of the group's swap limit for zram, a fixed 2 to 1 guess. The host watcher (tools/host-watch) now writes both words in its summary line and raises one finding when the peak is above that half: zram held <n> MB of memory at its peak since boot, more than the <half> MB the memory cap leaves for it …. The guess itself is unchanged.

boxd 0.4.71

Choose a tag to compare

@github-actions github-actions released this 22 Nov 00:10

A build server that cannot be reached is named, and does not hold a worktree it never ran.

  • A call whose sync fails because ssh got no answer now asks that server once and says box: <name> is unreachable (ssh got no answer); nothing was started (exit 255, as before). Until now only rsync's own lines and box: sync failed (rsync exit 255) were printed.
  • box --host <name> towards a server that is unreachable no longer leaves the worktree placed there: the placement is put back as it was before the call. Until now the next call without --host went to the same unreachable server although another one was free.
  • A worktree placed on a server that is unreachable stays placed (its box and caches may be there) and is told the way out: box: this worktree is placed on <name>; hosts that have room: <list>. Run it there with: box --host <first> ….

boxd 0.4.70

Choose a tag to compare

@github-actions github-actions released this 21 Nov 21:06

A new box says that its cache is shared.

  • When a call makes a box, a second line on standard error follows the one about the account: boxd-pod: box <box>: /cache is kept and shared by every box of this host (package caches, XDG_CACHE_HOME=/cache/xdg): a job finds there what earlier jobs put. A browser that one job installed under the cache made a test in a later box stop skipping and hang; the skill said so, the box did not.

boxd 0.4.69

Choose a tag to compare

@github-actions github-actions released this 21 Nov 18:45

A build host that reboots comes back by itself.

  • After a reboot every pod command could fail with current system boot ID differs from cached boot ID; an unhandled reboot has occurred and boxd-pod: podman ps failed. The build account has no runtime folder, so podman keeps its run state under the work area, on disk, and refuses to work after a reboot until two folders there are gone. The once-per-boot step of the build sshd (boxd-pod clear boot) now removes exactly those two folders before it clears the boxes.
  • Under runit the build sshd computed the memory cap before the zram service had switched its swap on, so the half of that swap that counts as memory was not taken off and the cap came out too high (about 2 GB on a 16 GB host). It now waits up to 30 seconds for the zram swap when the zramen service is enabled, and says so when it never comes. Under systemd the unit is ordered after zramswap.service.

boxd 0.4.68

Choose a tag to compare

@github-actions github-actions released this 21 Nov 09:04
  • A call made with --wait that had to wait now says when it goes on: box: room on <host> after waiting <time>; starting, on standard error, right before the command is handed to the host. Until now the caller saw box: waiting for room, ... and then nothing until the command's own output, and could not tell how long the wait had been. The time counts from the first wait; if the host refuses again (a lost race) and the call waits again, the line comes again before the next attempt. A call that never waited prints nothing new. Client only.

boxd 0.4.67

Choose a tag to compare

@github-actions github-actions released this 21 Nov 02:24
  • box pull <path>... --to <folder>: the pulled paths land under that folder instead of the worktree, each keeping its relative form (box pull out/report.txt --to /tmp/results gives /tmp/results/out/report.txt). Before, results always landed inside the calling worktree and had to be moved out by hand. --to without a folder is refused (exit 2). Without --to nothing changes. No server change.

boxd 0.4.66

Choose a tag to compare

@github-actions github-actions released this 21 Nov 00:08
  • New option --alive <path> for a detached run (box --detach --alive <path> ...): a file or folder of the worktree the run writes its progress to. Its newest change counts as output, so quiet= of box log --state and the silence line of box runs follow the job's own records. Before, a job that wrote only to its own files was reported as silent for hours while it worked. The path is relative, with no .. part; a folder is read at most 4 levels down and links are never followed; a missing path changes nothing. The server keeps the path in runs/<run>.alive and removes it with the run. The silence line of box runs now names the option. With --shards every shard's run gets it.

boxd 0.4.65

Choose a tag to compare

@github-actions github-actions released this 20 Nov 21:01
  • New word box pull <path>... [--host <name>] [--pod <name>]: copies paths of this worktree's box (its files under /job/src on the build server) back into the working folder and runs nothing in the box, so a stopped box, also a stopped virtual machine, stays stopped and needs no room on the host. Before, fetching results later meant box --no-sync --pull <path> true, which started a stopped machine for every read. A leading / or a .. in a path is refused (exit 2); a path that is not in the box prints box: nothing at <path> in the box, nothing pulled. The hints after a detached run name the new form. Because pull is now a box word, a program of that name run in the box needs -- before it. No server change.