carbon-tagger

Carbon-tagger

graphite protocol extension

Standard graphite protocol (proto1)

string.based.metric[...] value unix_timestamp

"string", "based", "metric" being the nodes of the metric, and metric names usually being unorganized, unstandardized and lacking information.

Carbon-tagger implements metrics 2.0 which aims to make metrics entirely self describing, structured, and standardized.

It does this by using an axtended graphite protocol (backwards compatible):

• nodes can be old-style values, or new-style key=val or key_is_val tag pairs. (Some versions of graphite-web/graphite-api struggle with equals signs, so the latter format is recommended)
• if there's a "=" or "is" in one or more of the nodes, we'll try to parse as proto2 and add it to the index if below conditions are met.
• there must be a tag pair with unit as tag key.
• there must be at least one other tag.
• you can freely choose the order of the nodes for every metric, but when you change the order, you change the metric key.
• old-style nodes (i.e. not "key=val" or key_is_val format) within a proto2 metric implicitly get an "nX" tag key where X is the node position in the string, starting from 1.

You'll probably want to follow the metrics naming conventions, specifically apply the correct units

So this defines a metric as a set of tags, and while we're at it, it also specifies the metric_id that will be used by the current carbon and related tools. carbon-tagger will maintain a database of metrics and their tags, and pass it on (unaltered) to a daemon like carbon-cache or carbon-relay. So the protocol is fully backwards compatible.

This is metrics 2.0 but

• using dots as delimiters until we can fix graphite.
• you can use units like "Mbps" or "Errps" to mean "Mb/s" and "Err/s". Graphite treats slashes as delimiters. Carbon-tagger will set the proper unit tag.

indexing

• Indexes metrics 2.0 full (_id and tag)
• legacy metrics, just the _id, so you can search for it. (empty tags property) it's up to a tool like graph-explorer to create or update documents for legacy metrics with tags enabled.

how does this affect the rest of my stack?

• carbon-relay, carbon-cache: unaffected, they receive the same data as usual, the identifiers just look a little different.
• graphite-web: unaffected. metrics 2.0 still show up in the tree of the graphite composer, although that's an inferior UI model that I want to phase out. ideally, dashboards leverage the tag database, like graph-explorer does.
• aggregators like statsd need proto2 support. statsdaemon is a drop-in statsd replacement that for metrics in metrics2.0 format correctly represents its aggregations and statistical summaries by updating the key/value pairs of metric names, as opposed to the traditional prefix/suffix "features" which leave metrics even vaguer than they were.

why do you process the same metrics every time they are submitted?

to have a realtime database. you could get all metricnames later and process them offline, which lowers resource usage but has higher delays

internal metrics

are in proto2 format and are submitted to a carbon endpoint (typically your relay) they are also available on the http address at /debug/vars2

performance

currently, not very optimized at all! but it's probably speedy enough, and there's a big buffer that smoothens the effect of new metrics

• I reach about 15k metrics/s processing speed, even when the temp buffer is full and it's syncing to ES. in fact, i don't see a discernable difference between buffer full (unblocked) and buffer full (blocked) probably carbon-cache (whisper) was being the bottleneck?

• space used: 176B/metric (21M for 125k metrics, twice that if we'd enable indexing/analyzing)

TODO

• make sure the bulk thing uses 'create' and doesn't update/replace the doc every time
• it seems like ES doesn't contain all metrics (on 2M unique inserts, ES' count is 1889300)
• better mapping, _source, type analyzing?
• don't store legacy metrics with empty tags, if already exists. now it will temporarily break graph-explorer's structured metrics (until the latter indexes again)

future optimisations

• populate an ES cache from disk on program start
• forward_lines channel buffering, so that runtime doesn't have to switch Gs all the time?
• GOMAXPROCS

if/when ES performance becomes an issue, consider:

• edge-ngrams/reverse edge-ngrams to do it at index time
• use prefix match with regular/reverse fields
• query_string maybe

building

if you already have a working Go setup, adjust accordingly:

mkdir -p ~/go/
export GOPATH=~/go
go get github.com/Vimeo/carbon-tagger
go get github.com/mjibson/party
go build github.com/Vimeo/carbon-tagger

installation

• just copy the carbon-tagger binary and run it (TODO: initscripts)
• install elasticsearch and run it (super easy, see http://www.elasticsearch.org/guide/reference/setup/installation/, just set a unique cluster name)
• ./recreate_index.sh