Julia nightlies broken? What can we do to help?

[catc...@bromberger.com]

Forgive me for stating what's probably painfully obvious to folks here, but it appears that julia nightlies have been failing for a couple of days and as a result, Travis builds are broken as well. Is there an ETA or workaround? Anything folks can do to help?

[Elliot Saba]

Hey there help is, of course, always welcome!  The nightlies are building (as you can see by looking at this page; feel free to ignore the status of the "package_cxx" jobs; those are special and are failing because of some LLVM stuff) and the last couple of Travis builds have either succeeded, failed because of a download problem (looks to have affected only a single build) or ran into a segfault.

Of all the problems mentioned above, the only one we can really get help with is tracking down that segfault.  On this build, you can see there was a segfault while running the cmdlineargs test.  This particular crash seems to be intermittent which can make it difficult to track down, but if you are so inclined, you could try running that test 100 times to see if it ever crashes, and if it does, you could attempt to catch it inside of a gdb session and poke around looking for what went wrong.  This could, of course, quickly spiral out of your knowledge or expertise of Julia, but could be a fun project for anyone looking to see what the internals of Julia look like.

That being said, it's entirely possible that the conditions necessary to trigger the segfault are OSX-specific, or only happen when there are 5 workers, or only happens 1 out of every 1000 runs, etc.... This is one reason these segfaults are difficult to pin down.

-E

[catc...@bromberger.com]

Ah. I think the confusion is mine - Travis appears to be failing because the change that made mathconsts irrational broke Distributions. I misinterpreted the CI build failure messages.

[Jameson Nash]

@Keno, that backtrace looks suspiciously similar to one you fixed earlier this spring.

In particular, I believe this is emitting a stringConst during recursive codegen evaluation. Unlike the earlier case, it would not be caught by jl_in_inference being true, since it is being triggered by running over the gc finalizer list rather than through recursive inference / codegen interleaving.

(also, perhaps, ref #5502)

[catc...@bromberger.com]

Jameson,

Could this be the cause of https://groups.google.com/d/msg/julia-users/CQ475y-mM7Y/ukAdazNObmcJ ? That is, the error message DimensionMismatch("") seems to imply that there's a string somewhere.

[Keno Fischer]

Can we delay running finalizers until after codegen is done?