Sergei Lebedev

IIRC there were some cross-platform compatibility issues related to `%z`. I'll have a look into this over the weekend.

**tl;dr** I think `":"` is the problem. Fact 1: ISO8601 [allows](https://en.wikipedia.org/wiki/ISO_8601#Time_offsets_from_UTC) UTC offset to be formatted with `":"`. Fact 2: `strptime` [Linux man page](http://man7.org/linux/man-pages/man3/strptime.3.html) claims to implement `%z` in a...

The [libc version](http://opensource.apple.com/source/Libc/Libc-1044.1.2/stdtime/FreeBSD/strptime.c) used in BSD and OS X doesn't work with the nonstandard `tm_gmtoff` field and instead directly subtracts the offset from `tm_hour` and `tm_min`. Linux does [exactly the...

Another solution would be to switch to Hadoop 3.X which [supports](https://hadoop.apache.org/docs/r3.1.1/hadoop-yarn/hadoop-yarn-site/UsingGpus.html) Nvidia GPUs natively.

I have been considering this, but I am afraid it is not straightforward: * dask-tensorflow uses hardcoded port range [2222, ....) and assumes that all of the ports are free....

@dubiousdavid so what was the problem with the above code?

My bad, I was referring to your code: is the current implementation of `listen!` causing the problem or is it the way you use `listen!`?

I think we can have a proper type for arrays in JAX, and I have a draft commit doing that internally. Trees, however, are harder, because their type is fundamentally...

Woah, looks like a bug to me. Interestingly, it is not present in `marisa_trie.Trie`: ```python >>> marisa_trie.Trie(["foo", "foo"]).keys() ['foo'] >>> marisa_trie.BytesTrie([("foo", b"1"), ("foo", b"2")]).keys() ['foo', 'foo'] ```

In theory you can do fuzzy matching with tries via branching search, but a more efficient way might be to use a [Levenshtein automata](http://blog.notdot.net/2010/07/Damn-Cool-Algorithms-Levenshtein-Automata). I've never used it in Python,...