Vladislav,
1 in 1000 operations
Currently, no such option exists. Tracing Scopes can only be enabled or
disabled cluster-wide.
However, even measuring the performance of a single operation using
Tracing cannot currently be considered representative.
The current Tracing implementation is not as lightweight as it might
appear. For example, SQL tracing benchmarks have shown performance
degradation of up to 50%, depending on the cache configuration.
The benchmark results are still available in the Apache Ignite Slack
|#tracing| channel, together with a discussion in which the authors of
the tracing implementation confirmed that similar performance
degradation was observed for other Tracing Scopes as well.
On 7/14/26 10:56, Vladislav Pyatkov wrote:
Mikhail,
The problem is that Tracing itself affects performance.
It looks like a bug, but it's not inherent to the tracing system. How could
tracing affect performance if it doesn't catch or intercept anything — for
example, only 1 in 1000 operations?
On Tue, Jul 14, 2026, 10:35 AM Mikhail Petrov<[email protected]> wrote:
Kirill,
When analyzing performance, it is better to have more tools rather than
fewer
The problem is that Tracing itself affects performance. As a result, in
its current state it is hardly a tool for answering "performance"
questions. I believe that Performance Statistics is a better fit here,
as it is less intrusive.
Where Tracing does shines - it is in visualizing distributed Ignite
processes in a user-friendly way and showing how the duration of
individual stages correlate.
However:
1. As I mentioned earlier, Ignite Tracing handles asynchronous
operations poorly, making the resulting traces difficult to interpret.
2. We are not aware of any examples of end users relying on this
mentioned visualization (see the related discussion on the user mailing
list).
How difficult is it to maintain?
The more important question is whether we want to maintain it at all. In
my opinion, the answer is no.
On 7/13/26 13:25, ткаленко кирилл wrote:
Hi Nikolai, I see what you mean.
I believe tracing is a pretty useful mechanism - provided it’s set up
and used correctly - and the alternatives that were proposed don't really
cover the same ground.
How difficult is it to maintain?
How much does it complicate development? I’d assume it shouldn't get in
the way too much.
Vladislav Pyatkov
