Go faster, Travis CI! #359
Conversation
For tests, we really don't need to do much more than build the compiler and run the tests --- Cabal shenanigans shouldn't be really necessary, so we can just use Stack here, which should handle caching much better. The `builtin-arrays' builds are currently broken, except on Dargent branches --- when those are merged, these builds should be flipped back to required. * Cogent ext2fs builds on Travis. * BilbyFs doesn't currently compile; it really should...
While we're here, bump resolvers.
|
|
||
| extra-package-dbs: [] | ||
| extra-deps: | ||
| - sbv-7.12 # for the `array' branches |
There was a problem hiding this comment.
A question: on my branch, I use sbv > 8. Which version of sbv can I have via extra-deps?
There was a problem hiding this comment.
This is only to get the pins right given the current builtin-arrays code. In theory, you can have whatever version you like. I believe sbv-8.x works fine, although I vaguely recall running into issues with crackNum, a sbv dependency, on GHC 8.2; I'd need to test that again to check, and I'd want to make sure stuff still compiles happily in general anyway.
|
@zilinc wrote:
We could test those. However, I believe it's more valuable to instead test the flag combinations (or to instead turn on more flags by default). e.g., how do recursive-types interact with Docgent or the Haskell backend? This would expand our test matrix massively, unfortunately. I propose we strip back supported versions to the latest on each of the GHC-x.y branches, unless there's good reason to maintain testing against the point-releases. Given we're also looking at GHC-8.8 support, perhaps this makes a good opportunity to also drop support for 8.2.x? |
jashank
force-pushed
the
jashank/2020/02/27/zoom-zoom-ci
branch
from
bda0cd5
to
9ee1060
Compare
* With +docgent, we use Pandoc, which won't Haddock in Stack. * With +haskell-backend, we get the haskell-lexer, which segfaults Haddock.
jashank
force-pushed
the
jashank/2020/02/27/zoom-zoom-ci
branch
from
9ee1060
to
329cd7e
Compare
Currently, CI cycle times are awful --- and really variable. The longest recent build I've seen is 9h45m21s, and most builds hover between 5h and 9h. Occasionally, the caches will smile on us and bestow a 45-minute-or-less cycle time, and even that's pretty miraculous.
After lots of experimenting, cajoling, and swearing, I think I've got CI cycle times down to about a quarter-hour or less, consistently.
Along the way, I've flensed out a lot of complexity in the build --- for tests, we really don't need to do much more than build the compiler and run the tests --- Cabal shenanigans shouldn't be really necessary, so we can just use Stack here, which is much more amenable to caching.
builtin-arrays-enabled builds are currently broken, except on Dargent branches --- when those are merged, these builds should be flipped back to required.A build that went OK, I guess:
Some builds that did not go well: