Hacker News Reader

Comment by yzmtf2008 | Hacker News Reader

yzmtf2008·

... which in turn resulted in TensorFlow not being able to be used in clusters for a long, long time: https://github.com/tensorflow/tensorflow/issues/110

View on HN

Which was the point. That was a long, long list of TensorFlow bugs that were avoided.

I just want to add to that, just to offer a perspective that people not concerned with the upkeep of software always realize: In the general case, when you keep software around for long some bugs will linger. But as long as the issues are well known that's not necessarily a bad thing.

Sometimes you have to work around some corner case or performance anomaly in another layer. You don't want to apply fixes to underlying software willy-nilly without first re-evaluating what you did.

So sometimes stable distributions will avoid fixing bugs for good reason. (That doesn't go for security issues, of course.) And if that's not what you want most distributions have a more suitable branch for you.

> most distributions have a more suitable branch for you

Exactly. CentOS has Fedora, Debian has sid/ubuntu/derivatives, Slackware has -current. If you don't require a stable platform (and most developers don't) then you're welcome to join people on the unstable side of things, making the next stable side of things.

The problem is most of the time developers working with clusters do not have control of which distribution to use. TensorFlow was considered stable (or at least stable enough for research) for a long time before it can be used with glibc 2.12.

That's odd, I was able to compile both bazel and tensorflow from source since early spring 2016 on RHEL 6.5 (not that it was straight forward ...), so it should work for centOS as well.

It certainly wasn't straightforward and took us two weeks to get right :(

Reply on news.ycombinator.com