go-graphite / carbonapi Goto Github PK

At the moment, each metric for a given target is fetched in parallel, but the targets themselves are processed serially. For queries with a large number of targets, this can can unnecessary slow down.

The function mostDeviant is missing/not implemented. The implementation of this function in the original graphite seems to be very slow.

If you specify a missing metric as part of a larger formula, the query will fail in carbonapi but succeed in Graphite.

A simple example of this is:

data.metric1  # Exists
data.metric2  # Does not exists

diffSeries(data.metric1, data.metric2) # Fails on carbonapi
diffSeries(data.metric1, transformNull(data.metric2, 0)) # Also fails on carbonapi

I would expect the result to be data.metric1 since I'd assume that data.metric1 - [undefined] would equal data.metric1.

A valid use case in this situation is as followed:

# Metric layout
page.[page_name].requests
page.[page_name].errors

# The number of total successful pages served (number of requests - number of errors)
# If we automate this in any form we may hit a page that has never had an error recorded.
diffSeries(page.index_html.requests, page.index_html.errors) 

If I send a request like this for a valid metric:
/info?target=metric_name
I always get the response:
empty target

If I sent the request like this:
/info/?target=metric_name
It's working.

Many graphing libraries use autogenerated jsonp callback names. This breaking caching heavily, since every request URI is now different.

When aggregating in a fixed time range with from/until, e.g., xxxxxx0 to xxxxxx84, aggregate by 5 seconds, there is sometimes an extra or a missing value, the array length is variable by 1.

Apparently, in a render url, the rawData flag, sometimes written as rawData=1 isn't supported. We should map it onto format=raw.

As mentioned in FB Gorilla paper add support for sorting/filtering metrics against a test metric using PPMCC

Work in progress on https://github.com/dgryski/carbonapi/tree/graph-support

Figure out the full set. Figure out which subset we want to support. Hope we don't have to fully break expression parsing in the meantime.

On carbonzipper restart, we will have a number of failed requests (~1 / cached idle connection). We don't want to add the results of these queries to our request or find cache.

https://github.com/dgryski/go-onlinestats/blob/master/tukey.go

Name parsing code is too limited

Sometimes the browser issues a HTTP request of the type OPTIONS when it's cross domain. I know a page that broke when using carbonapi because of this. It worked normally with the python version.

( Wikipedia description: Kolmogorov–Smirnov Test )

This would allow us to test, for example, a metric against itself in the past to see if the distribution has changed. Or if two metrics have similar distributions.

Naming might be among the hardest to implement.
My votes are:

• kolmogorovSmirnovTest2( seriesList, seriesList )
• ksTest2 (as the commonly used alias)

(I dislike the camelCasing standard here - but it maintains consistency)

"2" here because the 1-sample version, which essentially compares against a reference (usually normal) distribution, could also be implemented .. but there seem to be better tests for that.

It's getting unwieldy.

via @atomicstack

The response to an OPTIONS request doesn't include the CORS headers that an preflight AJAX request expects[1] which makes the browser to refuse to perform the request.

[1] https://developer.mozilla.org/en-US/docs/Web/HTTP/Access_control_CORS#Preflighted_requests

Graphite munges timestamps to minutely resolution to increase the chance of a cache hit. It also stores the data returned by the stores so future queries on the same data can also be pulled from the cache.

Decide which of these we want to implement.

There are people that would like to use with carbonapi the holtWinters functions present in the python version

What license is this and carbonzipper under?

It would be nice to have it. Documentation here:
https://graphite.readthedocs.org/en/0.9.10/functions.html

In addition to code.google.com shutting down, plotinum has moved to being part of the gonum project.

This move also comes with a substantial number of breaking API changes.

example request: https://graphite-api.[redacted]/render?target=aliasByNode(stacked(sys.*[redacted]*.[redacted].connections)%2C%201)&from=1439439654&until=1439482974&format=json&maxDataPoints=640

It turns out people have very poor graphite queries that fail to parse. Figuring out where the missing comma needs to go is tedious at the moment. Make this easier.

Or is there a way in there right now to multiply two series together? Can't seem to figure out a good way in grafana with the current api, I think multiplySeries() would do the job.

scale(foo.bar.baz, 1e-2)

http://graphite.readthedocs.org/en/latest/functions.html#graphite.render.functions.minSeries

Need to figure out "only minutely resolution" for queries, <= vs <, etc.

Because we maintain lots of cached http connections to carbonzipper, when that service is restarted we have a large number of (now invalid) sockets which cause failures during subsequent requests.

Work around is to restart carbonapi when carbonzipper is upgraded.

http://graphite.readthedocs.org/en/latest/functions.html#graphite.render.functions.hitcount

[This issue serves as a reminder / todo of interval's inteneded use.]

interval was conceived as a mechanism where the quantile threshold can be computed on a (leading IIRC) subset to enable this funciton to more efficiently process large seriesList over large intervals.

@dgryski and I discussed that we could move it to the end as an optional parameter.

Question: preserve graphite bug compatibility or not?

time, ip, request, response size, ... ?

I try to implement functions on demand.

These functions have been demanded.

Graphite handles both, but Go's default muxer doesn't.

Right now we just return an empty metric list, instead a 500 or a json blob with a useful error message.

This will throttle bulk-style requests, but not adversely affect smaller API calls from proceding.

Requires a bit more thinking to figure out if this is a problem and if this is the best solution.

It is documented here:

http://graphite.readthedocs.org/en/latest/render_api.html

and it seems to be used by grafana:

http://docs.grafana.org/installation/performance/

Note that the breakage might actually be in carbonzipper or carbonserver.

Our internalString grammar is too simple.

http://graphite.readthedocs.org/en/latest/functions.html#graphite.render.functions.sumSeriesWithWildcards

The graphite front-end stores responses in memcache, meaning that subsequent requests for the same metric are much faster and don't hit the stores. This is important when lots of people are looking at the same graph or fetching the same metric.

An in-memory version is fine for starters, but we should move to externalize it so it can be shared across carbonapi instances.