Refs: #4419
Refs: #4453
Refs: nodejs/node#65826
While looking at the disk pressure on the built-in node i measured what one PR run writes in console logs. node-test-pull-request 77023 (a normal, complete run) is 84 builds and 482 MB of console text; 475 MB of that is the 35 configurations that compile (node-test-commit-linux, -plinux, -linuxone, -arm, -arm-debug, -aix, -osx, -smartos, -linux-containered), 12-22 MB each, and in every one of them everything before TAP version 13 is compiler command lines - e.g. nodes=rhel8-x64 is 15.2 MB with 14.7 MB before the tests start. For comparison a node-test-binary-windows-js-suites console is 0.17 MB and node-compile-windows ~1 MB. At ~15 PR runs a day that's on the order of 7 GB/day of console from PR testing alone, before the daily jobs, which lines up with the compressor -mtime juggling in #4419.
Two things produce it. Most jobs run plain make run-ci -j N, and node's Makefile has been forwarding V=0 to the gyp makefile, which treats any non-empty V as verbose; nodejs/node#65826 fixes that, so once it lands (and reaches the release lines) those consoles should drop to well under 1 MB each with no change here. node-test-commit-linux is the exception: its job config runs make run-ci -j 2 V=1 explicitly (visible at the top of any of its consoles), so its 10 configurations (~155 MB per PR run) stay verbose until that V=1 is removed from the job.
The ask is to drop V=1 from node-test-commit-linux's shell step (and anywhere else it's set that i haven't spotted). If full command lines are wanted somewhere for toolchain debugging, keeping V=1 on a single configuration would preserve that at a tenth of the cost.
Refs: #4419
Refs: #4453
Refs: nodejs/node#65826
While looking at the disk pressure on the built-in node i measured what one PR run writes in console logs.
node-test-pull-request77023 (a normal, complete run) is 84 builds and 482 MB of console text; 475 MB of that is the 35 configurations that compile (node-test-commit-linux,-plinux,-linuxone,-arm,-arm-debug,-aix,-osx,-smartos,-linux-containered), 12-22 MB each, and in every one of them everything beforeTAP version 13is compiler command lines - e.g.nodes=rhel8-x64is 15.2 MB with 14.7 MB before the tests start. For comparison anode-test-binary-windows-js-suitesconsole is 0.17 MB andnode-compile-windows~1 MB. At ~15 PR runs a day that's on the order of 7 GB/day of console from PR testing alone, before the daily jobs, which lines up with the compressor-mtimejuggling in #4419.Two things produce it. Most jobs run plain
make run-ci -j N, and node's Makefile has been forwardingV=0to the gyp makefile, which treats any non-emptyVas verbose; nodejs/node#65826 fixes that, so once it lands (and reaches the release lines) those consoles should drop to well under 1 MB each with no change here.node-test-commit-linuxis the exception: its job config runsmake run-ci -j 2 V=1explicitly (visible at the top of any of its consoles), so its 10 configurations (~155 MB per PR run) stay verbose until thatV=1is removed from the job.The ask is to drop
V=1fromnode-test-commit-linux's shell step (and anywhere else it's set that i haven't spotted). If full command lines are wanted somewhere for toolchain debugging, keepingV=1on a single configuration would preserve that at a tenth of the cost.