RFC: Would a config fetch.retryCount make sense?

Hi,

we occasionally see "The remote end hung up unexpectedly" (pkt-line.c:265) on our `git fetch` calls (most noticeably in our automations). I expect random network glitches to be the cause.

In some places we added a basic retry mechanism and I was wondering if this could be a useful feature for Git itself.

E.g. a Git config such as "fetch.retryCount" or something. Or is there something like this in Git already and I missed it?

Thanks, Lars

On 6/1/2017 8:48 AM, Lars Schneider wrote:

Having a configurable retry mechanism makes sense especially if it allows continuing an in-progress download rather than aborting and trying over. I would make it off by default so that any existing higher level retry mechanism doesn't trigger a retry storm if the problem isn't a transient network glitch.

Internally we use a tool (https://github.com/Microsoft/GVFS/tree/master/GVFS/FastFetch) to perform fetch for our build machines. It has several advantages including retries when downloading pack files.

It's biggest advantage is that it uses multiple threads to parallelize the entire fetch and checkout operation from end to end (ie the download happens in parallel as well as checkout happening in parallel with the download) which makes it take a fraction of the overall time.

When time permits, I hope to bring some of these enhancements over into git itself.

Agreed.

That's a "drop-in" replacement for "git fetch"?! I looked a bit through the "git fetch" code and retry (especially with continuing in-progress downloads) looks like a bigger change than I expected because of the current "die() in case of error" implementation.

Interesting. Do you observe noticeable speed improvements with fetch delta updates, too? This is usually fast enough for us.

The people I work with usually complain that the "clone operation" is slow. The reason is that they clone over and over again to get a "clean checkout". I try to explain to them in that case that every machine should clone only once and that there are way more efficient ways to get a clean checkout.

That would be great!

- Lars

On 6/5/2017 8:04 AM, Lars Schneider wrote:

No, not a drop in replacement. We only use this on build machines which don't need history so it only pulls down the tip commit on the initial clone. This is a big win on large repos with a lot of history but not so great for a developer machines where history may be desired.

Since we have our build machines setup to use it for the clone, we kept using it for delta updates. When deltas get large (and with thousands of developers pushing that can happen pretty quickly) it is still a nice perf win.

On Thu, Jun 1, 2017 at 5:48 AM, Lars Schneider <larsxschneider@gmail.com> wrote:

There is 665b35eccd (submodule--helper: initial clone learns retry logic, 2016-06-09) but that is for submodules and only the initial clone.

I tried searching the mailing list archive if it was discussed for fetch before (I am sure it was), but could not find a good hint to link at.

IIRC one major concern was:
* When a human operates git-fetch, then they want to have fast feedback.
  The failure may be non-transient, for example when I forgot to up the
  wifi connection. Then the human can inspect and fix the root cause.
  (Assumption in human workflow: these non transient errors happen more
  often than the occasional fetch error due to network glitches.)

For automation I would expect that the retry logic is actually beneficial, such that you would want to have a command line options such as "git fetch --retries=5 --delay-between-retries=10s".

There are already retries in other places. :) Cf. f4ab4f3ab1 (lock_packed_refs(): allow retries when acquiring the packed-refs lock, 2015-05-11), which solves the need of github on the serverside, when they have a very active repo that multiple people push to at the same time. (to different branches. I believe that forks are internally handled as the same repo, just with different namespaces. So if there are 1000 forks of linux.git you see a lot of pushes to the "same" repo)

I like it.

Thanks, Stefan