On Mon, Dec 11, 2017 at 9:53 AM, James Peach <jor...@gmail.com> wrote:
> > > On Dec 11, 2017, at 8:55 AM, Zhitao Li <zhitaoli...@gmail.com> wrote: > > > > Hi, > > > > We are running tasks which has very large docker images and tasks which > use > > much smaller images in our clusters, Therefore, we expect to see > occasional > > violation of --executor_registration_timeout for tasks which has > uncached > > large docker images. > > > > I wonder whether we can introduce some executor specific parameter to > make > > this customizable per executor, instead of one single value per agent. > > Sounds like the registration timeout should not start until all the > required images have been staged? > +1, it has come up before that the registration timeout shouldn't include fetching, that was largely an accident as the code evolved IIUC. The fetching of artifacts / docker images should probably have its own timeout / minimum speed / etc specification. > J