Sorry, one more thing occurred to me. Can the names of the time-based
metrics include their units? That makes it much easier for people
consuming the metrics to interpret them correctly.
For example, offset-commit-max-time would become offset-commit-max-time-ms
or
FWIW, I agree that time metrics have been very useful in the past. The
reasoning around perf overhead seems reasonable as well. Can we agree on a
subset of time metrics that we feel would be super useful for debugging?
On Tue, Sep 12, 2017 at 6:08 PM, Roger Hoover
wrote:
Thanks, Ewen.
I agree with you on the overhead of measuring time for SMTs and
converters. I'd still argue for keeping other metrics like flush time b/c
even small batches should still be small overhead compared to writing to a
sink.
On Tue, Sep 12, 2017 at 3:06 PM, Ewen Cheslack-Postava
Regarding the existing rebalance metrics under
"kafka.connect:type=connect-coordinator-metrics", I think we should just
plan on reusing them rather than duplicating them.
On Tue, Sep 12, 2017 at 5:06 PM, Ewen Cheslack-Postava
wrote:
> Requests are generally substantial
Requests are generally substantial batches of data, you are not guaranteed
that for the processing batches both because source connectors can hand you
batches of whatever size they want and consumer's max.poll.records can be
overridden.
Both SMTs and converters are a concern because they can both
Ewen,
I don't know the details of the perf concern. How is it that the Kafka
broker can keep latency stats per request without suffering too much
performance? Maybe SMTs are the only concern b/c they are per-message. If
so, let's remove those and keep timing info for everything else like
On Tue, Sep 12, 2017 at 10:55 AM, Gwen Shapira wrote:
> Ewen, you gave a nice talk at Kafka Summit where you warned about the
> danger of SMTs that slow down the data pipe. If we don't provide the time
> metrics, how will users know when their SMTs are causing performance
>
Okay, I think I've incorporated all feedback except for Gwen and Roger than
would like to have timing metrics. Given the deadline and Ewen's concern
about degraded performance, I think it's prudent to leave those out of this
KIP and proceed as is.
On Tue, Sep 12, 2017 at 12:48 PM, Randall Hauch
Ewen, you gave a nice talk at Kafka Summit where you warned about the
danger of SMTs that slow down the data pipe. If we don't provide the time
metrics, how will users know when their SMTs are causing performance issues?
Gwen
On Mon, Sep 11, 2017 at 7:50 PM Ewen Cheslack-Postava
On Tue, Sep 12, 2017 at 10:36 AM, Roger Hoover
wrote:
> Randall/Ewen,
>
> I think the timing info is still useful even if it's measured since the
> last rebalance. How else do you know where time is being spent?
>
I think Ewen's concern (correct me if I'm wrong) is that
Hi, James. I was mistaken about how the Kafka metrics are converted to
MBeans and attributes. The MBean is constructed from the group and tags,
and the metrics show up as attributes on the MBean. I'll update the KIP to
reflect this.
On Tue, Sep 12, 2017 at 1:43 AM, James Cheng
Randall/Ewen,
I think the timing info is still useful even if it's measured since the
last rebalance. How else do you know where time is being spent?
The use case for seeing the batch size is that you generally have two knobs
to configure - max batch size and max wait time. The batch size
Thanks for the KIP, Randall.
The KIP has one MBean per metric name. Can I suggest an alternate grouping?
kafka.connect:type=connector-metrics,connector=([-.\w]+)
connector-type
connector-class
connector-version
status
re: questions about additional metrics, I think we'll undoubtedly find more
that people want in practice, but as I mentioned earlier I think it's
better to add the ones we know we need and then fill out the rest as we
figure it out. So, e.g., batch size metrics sound like they could be
useful, but
On Mon, Sep 11, 2017 at 4:50 PM, Randall Hauch wrote:
> Thanks, Ewen. Comments inline below.
>
> On Mon, Sep 11, 2017 at 5:46 PM, Ewen Cheslack-Postava
> wrote:
>
> > Randall,
> >
> > A couple of questions:
> >
> > * Some metrics don't seem to have unique
Based on Roger and Ewen's feedback, I removed the aggregate metrics as they
would be difficult to make use of without extra work. This simplified
things a great deal, and I took the opportunity to reorganize the groups of
metrics. Also, based upon Ewen's concerns regarding measuring
Thanks, Ewen. Comments inline below.
On Mon, Sep 11, 2017 at 5:46 PM, Ewen Cheslack-Postava
wrote:
> Randall,
>
> A couple of questions:
>
> * Some metrics don't seem to have unique names? e.g.
> source-record-produce-rate and source-record-produce-total seem like they
> are
Randall,
A couple of questions:
* Some metrics don't seem to have unique names? e.g.
source-record-produce-rate and source-record-produce-total seem like they
are duplicated. Looks like maybe just an oversight that the second ones
should be changed from "produce" to "write".
* I think there's a
Thanks for reviewing. Responses inline below.
On Mon, Sep 11, 2017 at 11:22 AM, Roger Hoover
wrote:
> Randall,
>
> Thank you for the KIP. This should improve visibility greatly. I had a
> few questions/ideas for more metrics.
>
>
>1. What's the relationship between
Randall,
Thank you for the KIP. This should improve visibility greatly. I had a
few questions/ideas for more metrics.
1. What's the relationship between the worker state and the connector
status? Does the 'paused' status at the Connector level include the time
that worker is
Thanks, Gwen.
That's a great idea, so I've changed the KIP to add those metrics. I've
also made a few other changes:
1. The context of all metrics is limited to the activity within the
worker. This wasn't clear before, so I changed the motivation and metric
descriptions to explicitly
Thanks for the KIP, Randall. Those are badly needed!
Can we have two metrics with record rate per task? One before SMT and one
after?
We can have cases where we read 5000 rows from JDBC but write 5 to Kafka,
or read 5000 records from Kafka and write 5 due to filtering. I think its
important to
Hi everyone.
I've created a new KIP to add metrics to the Kafka Connect framework:
https://cwiki.apache.org/confluence/display/KAFKA/KIP-196%3A+Add+metrics+to+Kafka+Connect+framework
The KIP approval deadline is looming, so if you're interested in Kafka
Connect metrics please review and provide
23 matches
Mail list logo