On Wed, Jan 25, 2017 at 9:55 PM, Sam Overton <samover...@gmail.com> wrote: > Hello cassandra-dev, > > I would like to continue the momentum on improving Cassandra's tracing, > following Mick's excellent work on pluggable tracing and Zipkin support. > > There are a couple of areas we can improve that would make tracing an even > more > useful tool for cluster operators to diagnose ongoing issues. > > The control we currently have over tracing is coarse and somewhat > cumbersome. > Enabling tracing from the client for a specific query is fine for > application > developers, particularly in an environment where Zipkin is being used to > trace > all parts of the system and show an aggregated view. For an operator > investigating an issue however, this does not always give us the control > that we > need in order to obtain relevant data. We often need to diagnose an issue > without the possibility of making any changes in the client, and often > without > the prior knowledge of which queries at the application level are > experiencing > poor performance. > > Our only other instigator of tracing is nodetool settraceprobability which > only > affects a single node and gives us no control over precisely which queries > get > traced. In practise, it is very difficult to find the relevant queries that > we > want to investigate, so we have often resorted to bulk loading the traces > into > an external tool for analysis, and this seems sub-optimal when cassandra > could > reduce much of the friction. > > I have a few proposals to improve tracing that I'd like to throw out to > the mailing list to get feedback before I start implementing. > > 1. Include trace_probability as a CF level property, so sampled tracing can > be > enabled on a per-CF basis, cluster-wide, by changing the CF property. > https://issues.apache.org/jira/browse/CASSANDRA-13154 > > 2. Allow tracing at the CFS level. If we have a misbehaving host, then it > would > be useful to enable sampled tracing at the CFS layer on just that host so > that > we can investigate queries landing on that replica, rather than just queries > passing through as a coordinator as is currently possible. > https://issues.apache.org/jira/browse/CASSANDRA-13155 > > 3. Add an interface allowing for custom filters which can decide whether > tracing > should be enabled for a given query. This is a similar idea to > CASSANDRA-9193 > [1] but following the same pattern that we have for IAuthenticator, > IEndpointSnitch, ConfigurationLoader et al. where the intention is that > useful > default implementations are provided, but abstracted in such a way that > custom > implementations can be written for deployments where a specific type of > functionality is required. This would then allow solutions such as > CASSANDRA-11012 [2] without any specific support needing to be written in > Cassandra. > https://issues.apache.org/jira/browse/CASSANDRA-13156 > > Thanks for reading! > Regards, > > Sam > > > [1] https://issues.apache.org/jira/browse/CASSANDRA-9193 Facility to write > dynamic > code to selectively trigger trace or log for queries > > [2] https://issues.apache.org/jira/browse/CASSANDRA-11012 Allow tracing CQL > of a > specific client only, based on IP (range)
Not directly related, but to make (3) more useful it would also be great to be able to list currently executing queries. I've had multiple cases where read queries would just use all my slots and never finish and it was quite painful to discover what the query was exactly (slow query don't help if the query never finishes). -- Corentin Chary http://xf.iksaif.net
