This sounds good to me! On Mon, Feb 5, 2018 at 1:08 AM, Ning Wang <wangnin...@gmail.com> wrote:
> Yeah. That is an option too. In fact it was my first try: > https://github.com/twitter/heron/pull/2693 (just an initiative, not > completed, a count map should be used instead of a single total count) > > In most cases, I think both solutions should have the same result. A few > reasons I changed to a tmaster check: > - with tmaster, there is only one source of truth and tmaster is more > critical anyway. If the tmaster link is not healthy, stmgrs won't work > correctly: topology may have created replacement nodes but the disconnected > nodes could keep going by themselves. > - it is more straightforward. The logic is the same as the current one. One > the other side, if we use an array for all remote stmgrs, we could have a > smarter logic (which is good) but it could make stmgrs more complicated and > less straightforward (bad). I left the stmgr counters there so if in future > we decide to add this feature, it should be easy to add. There is a gap > between "errors from all" and "errors from a few" and this is not a > simple/quick question. > > > > > On Sun, Feb 4, 2018 at 6:48 PM, Sanjeev Kulkarni <sanjee...@gmail.com> > wrote: > > > I could't add comments to the document, thus am posting my comments to > the > > mailing list > > One more approach could be to do the current measurement as it is, but > > instead of leaving the quitting decision to the stmgtclient, have > > stmgrclientmgr do the decision. Thus everytime a stmgr client detects > > connection issues, inform that to stmgrclientmgr which keeps a map of > > peerstmgrid to error count. Thus it is able to decide things like am i > > seeing connection errors from all stmgrs or if only a few of them are > > having issues. Then it can take the decisions better. > > > > On Sat, Feb 3, 2018 at 8:11 PM, Ning Wang <wangnin...@gmail.com> wrote: > > > > > Hi, heron devs~ > > > > > > I think the current stream manager's quitting logic on connection > > failures > > > is problematic. We saw a few internal cases in Twitter that this logic > > > could cause extra issue. > > > > > > Here is a doc with more details: > > > > > > https://docs.google.com/document/d/1WHNc2NEp2gVL9ge2QVKp9t4Hpd4U9 > > > sAbzBqCu4-iDUM/edit# > > > > > > Comments and feedbacks are welcome! > > > > > > Thanks. > > > --ning > > > > > >