We won't proceed in the face of a missing table, we'll just crash. But it's still a bad experience for us and the end user; we have to guess that maybe a TimeoutException means a missing topic, and we also have to wait the N seconds (default 60) for the thing to "timeout".
On Tue, Oct 25, 2016 at 12:14:46AM -0700, Andy Chambers wrote: > You could just catch the exception but if this is per row, that is probably > prohibitively expensive. > > Doesn't the binlog get "create table" events? Wouldn't that be a better > time to create the topic? > > -- > Andy > > On Mon, Oct 24, 2016 at 2:32 PM, Ben Osheroff <b...@zendesk.com.invalid> > wrote: > > > Hiya! > > > > I've been trying to merge https://protect-us.mimecast.com/s/ANVVBZU83nznf9, > > which adds a much-requested feature of Maxwell, that of being able to > > have a topic-per-mysql-table. When we receive a row we programmatically > > generate the topic name, and the first thing we do is call > > `KafkaProducer#partitionsFor(topic)`, so that we know how to partition > > the data. > > > > The problem I'm running into is in trying to detect the case where a > > topic doesn't exist. If auto-creation is on, `partitionsFor()` seems to > > correctly auto-create the topic, but if auto-creation is off the > > behavior is kinda wonky; kafka goes into a metadata-fetch loop, logging > > > > "Error while fetching metadata with correlation id 573 > > :{topic=UNKNOWN_TOPIC_OR_PARTITION}" > > > > but then ultimately throwing me back a `TimeoutException` after 60 tries > > or so. > > > > I can rescue/rethrow the TimeoutException, but it seems like there might > > be a better way that I'm missing. Any ideas? I'd ideally just like a > > way to fail fast and clean when the topic doesn't exist (and > > auto-creation is off). > > > > Thanks, > > Ben Osheroff > > zendesk.com > > > > > > > >
signature.asc
Description: PGP signature
