CAT Graphite replacement lightning talk

Riemann + InfluxDB + Grafana

A more performant, easier to deploy and hopefully more scalable replacement for Graphite

Nick Chappell

https://github.com/nickchappell

[email protected]

Graphite

Where did it come from?Developed internally at Orbitz, then open-sourced

Why did it catch on?

Queryable HTTP API, which RRD didn’t (AFAIK, still doesn’t) have anything similar to

Once installed, easier to work with than RRD

Coolness factor!Lawl

GraphiteWhat is it?

Metrics receiver (carbon-cache)

Metrics relay (carbon-relay)

Metrics data point storage (Whisper)

Metrics data graphing (Graphite web)

Metrics data API (Graphite web)

Graphite architecture3 main components

whisper

graphite-web

Carbon takes in metrics

Whisper stores metrics

graphite-web retrieves metrics

carbon-cache/carbon-relay

API

Other tools in the graphing/metrics space

Old school toolsRRD for data storage

RRDtool for generating graphs

Cacti for managing dashboards of RRD

graphs

GraphiteWhat’s wrong with it?

A few different things, mainly with 2 parts of the stack

carbon-cache/carbon-relay whisper

graphite-web

graphite-webGraphite web is actually quite good and featureful (is that a word?) as an API

The built-in dashboard and graph builder isn’t very stylish, but is great for exploring and finding out what metrics you want to query via the API or graph in another dashboard

tool

What’s wrong with Whisper?TONS of files, 1 file per metric

Disk IO problems when running queries for lots of metrics !

graphite-web ends up having to touch lots and lots of files

Whisper derpage

This is your timeseries data on catnip

What’s wrong with carbon?The stock Python interpreter, CPython, specifically, the GIL

(global interpreter lock)

“Threading” doesn’t really happen because multiple threads are not allowed to execute native code at the same time

https://wiki.python.org/moin/GlobalInterpreterLock

Why? According to the Python folks, memory management is not thread safe

Your time series data is getting munged on somewhere in here

Does Graphite have any redeeming qualities?

The format that Graphite receives metrics in is dead-simple

metric_name value timestamp\n

foo.bar.baz 42 74857843

walle.ece.cecs.pdx.edu.cpu-system 423 74857843

Metrics formatsFor better or worse, Graphite’s format is the lingua franca of

the metrics world right now

Most everything that outputs metrics can output them in Graphite’s format

Metrics 2.0

Graphite’s format is too limited to store more data or more complex types of metrics

http://metrics20.org/

Metrics 2.0 is an informal proposal by Dieterbe, the monitoring/metrics dude at Vimeo

What can we do?

We can tune and scale Graphite by architecting things differently or we can replace some or all of

the components

Scaling GraphiteSome strategies include…

Split up the roles for Carbon into Carbon relays and Carbon caches

…decoupling components

Get a faster CPU

…throwing more hardware at it

Get faster disks (SSDs or set up RAM drives)

Put HAproxy in front of a few Carbon instances to spread the load around

To this end, set up multiple Carbon instances on the same machine, listening on different ports

and tie them to separate CPU cores

Will it work?Maybe. But you get a more complex setup

Once you’ve hit a ceiling in terms of how much disk IO you can provide, Whisper runs out of juice

If you have multiple carbon caches, do they write to the same Whisper data store?

Is it even possible to write to the same Whisper data store?

If not, how do we set up a graphite-web instance to query multiple Whispers?

If we can do that, how does each instance know which Whisper to query for what metric series?

What are other people doing?

Some people are trying to rewrite parts of the stack

https://github.com/graphite-ng/

http://grey-boundary.com/the-architecture-of-clustering-graphite/

Others have set up some pretty impressive (but complex) Graphite architectures:

http://grey-boundary.com/the-architecture-of-clustering-graphite/

What are the Graphite people doing?

The Graphite folks have 2 projects in the works to overcome some of its problems

megacarbon: replacement for Carbon, supposedly will perform better

ceres: replacement for Whisper, supposedly will perform better and natively allow writes from multiple carbon/

megacarbon instances

Will these help?Probably not.

megacarbon is a branch of the main carbon repository that is ahead of carbon’s master branch, but hasn’t yet been merged

back in and may never be

ceres hasn’t been touched since December 2013

For that matter, whisper hasn’t been touched since January 10th, 2014

Issues and pull requests keep piling up

One more thing….Graphite is a PITA to install/deploy

It’s somewhat easier now that carbon, whisper and graphite-web are available in pip

You still need to do a bunch of manual setup, though, and use apt/yum to install a bunch of

pre-requisite Python packages and set up Apache/Nginx with WSGI or Gunicorn to run graphite-web, which is actually a Django app

Before, you had to clone each component’s Git repo and check out a tagged release

OK, what next?

Outside of planet Graphite, other things have been happening in the monitoring/metrics/

logging space

Let’s take a step backFor a long time, “monitoring” was just this:

What a sleepless night will look like

And if it’s all green, things are OK, right?

Proactive vs reactive

I get alerts about disk space going below a certain threshold or load average going above a certain threshold

Can I get alerts about those things before they become problems?

Can I get alerted about disk usage when it may not be problem now, but is on its way up and soon will be critical?

“Traditional” monitoring tools

Nagios, Icinga, Sensu, Shinken, etc. are great at monitoring state changes

They’re not very good at gathering metrics or performance data, and they’re not built to look for trends in things over time

Things going from good (server is up, site is responding, database is running) to bad or bad to worse

They can gather performance stats, but the intervals are too long to be useful

Nagios/Icinga with plugins can generate reports about trends (how many services were critical in the past month) but it’s after the fact

3 newer monitoring tools

Logstash

Heka

Riemann

Logstash

Accepts logs, processes and stores them in Elasticsearch

A web app, Kibana, accesses the index data in Elasticsearch

Sound familiar?

There are tons of other ways to use it, though!

Heka

Newest monitoring tool of the bunchWritten in Go

https://github.com/mozilla-services/heka

Riemann http://riemann.io

What is an event?“Event stream processor” sounds really abstract

Instead of metrics (CPU usage on webserver2 at 10:45:09am) log lines and monitoring alerts (MySQL was down at 12:54am)

being thought of as separate things, why think of them as different variations of the same thing?

Logstash was one of the first tools to work with this assumption

Think about it: a user hitting a 503 error in your web app (that you find out via logs) may not trigger a Nagios alert, but don’t

you want to still know about it?

RiemannAn “event stream” processor

http://riemann.io

RiemannCan take in almost anything (log line, Graphite-format metric,

etc.)

Like Logstash/Heka, can process it and then send it elsewhere

Outputs include…

…IM protocols/services like IRC, Slack, Hipchat

…email

…a more traditional monitoring system like Nagios/Icinga

RiemannCan also output to Graphite and Logstash

I’ll talk about integrations with other tools later on

One advantage over carbon: Riemann has Debian and RPM packages, including init scripts!

Riemann as part of a Graphite replacement

Riemann can actually take Graphite-format metrics as an input

http://riemann.io/api/riemann.transport.graphite.html

OK, so can Riemann be used in the place of Carbon?

Yes. It can be instructed to listen on arbitrary TCP and UDP ports

RiemannWritten in Clojure, a functional programming language

Riemann being written in Clojure is kinda neat, but what makes it special is what Clojure runs on: the JVM

Clojure also brings one benefit that Graphite desperately needs: safe threading!

Unlike CPython, the JVM can actually run more than 1 thread at a time and can use more than 1 CPU

core

Clojure has software transactional memory and other tools for parallel/concurrent programming

Riemann events

Like Logstash, events in Riemann are pieces of text with multiple fields

{ :host riemann1.local, :service cpu-0.cpu-wait, :metric 3.399911, :tags collectd, :time 1405715017, :ttl 30 }

Other parts of the stackSo, we have Carbon replaced. What else do we need?

A replacement for Whispher

A replacement for the API component of graphite-web

A replacement for the web UI component of graphite-web

InfluxDB

InfluxDB

A fairly young project (about a year old)

A time series database written in Go

Uses LevelDB as the underlying datastore, but about to switch to RocksDB as a default (LevelDB, HyperLevelDB and LMDB can also be used)

RocksDB and HyperLevelDB are based on LevelDB

All 3 LevelDB variants compress data on disk

All 4 storage engines can compact storage engine files when getting rid of old metrics

InfluxDB specificsEven though it’s called a time series database, it isn’t

like a regular SQL database

A database in InfluxDB is just like a DB in a SQL system

A series in InfluxDB is like a table in SQL

A point in a series is like a row in a table

Points can have columns of values

Points in a series don’t all have to have the same columns, so InfluxDB is sort of schema-less

InfluxDB advantages over WhisperIt has a query language (usable via an HTTP API)

All of the underlying storage engines can perform better than Whisper

http://influxdb.com/blog/2014/06/20/leveldb_vs_rocksdb_vs_hyperleveldb_vs_lmdb_performance.html

Storage engine benchmarks:

InfluxDB is going to move to RocksDB as the default in the next version

They have Debian and RPM packages! (with init scripts too!)

Can InfluxDB replace Whisper?Yes!

One particular feature it has that Whisper does not is clustering

A group of InfluxDB nodes can communicate via a Raft-based protocol to coordinate writes and reads and split up data into

shards

Whisper is still single-instance only

InfluxDB ends up taking the place of Whisper and graphite-web

InfluxDB disadvantages

API and built-in functions are not as featureful as graphite-web, at least not yet

Data retention specifics and best practices are still being worked out

But, both of these can be solved with development work, and unlike the Graphite projects, InfluxDB is being actively

developed!

Riemann + InfluxDBGuess what Riemann can write data out to?

InfluxDB data partitioningData partitioning is deciding how to organize your data (1

series per host? 1 series per metric? 1 series per hostname + metric name combo?)

InfluxDB works best with large numbers of series with fewer columns in each one

Why? Points are indexed by time, not by any other columns.

Arbitrary column indexes are going to be added in the future, though

Having only 1 hostname+metric combo’s set of data in a series means InfluxDB only has to do fast indexed lookups by time,

not fast indexed lookups by time, then slower non-index lookups of other columns through several hosts’ and metrics’

worth of data

InfluxDB data partitioning

Time Name Host Metric Service32141234 cpu web0

178 cpu

32141235 disk_io web02

98844 disk_io32141236 load db1 5 load32141237 eth0_in ldap0

35875 eth0_in

Time Name Host Metric Service32141234 cpu web01 78 cpu32141235 cpu web01 45 cpu32141236 cpu web01 38 cpu32141237 cpu web01 92 cpu

Time Name Host Metric Service32141234 disk_io web01 87323 disk_io32141235 disk_io web01 98844 disk_io32141236 disk_io web01 9233 disk_io32141237 disk_io web01 93262 disk_io

Bad: only using 1 series in a DB

Good: Using multiple series in a DB

Time Name Host Metric Service32141234 cpu web02 78 cpu32141235 cpu web02 45 cpu32141236 cpu web02 38 cpu32141237 cpu web02 92 cpu

Time Name Host Metric Service32141234 disk_io web02 87323 disk_io32141235 disk_io web02 98844 disk_io32141236 disk_io web02 9233 disk_io32141237 disk_io web02 93262 disk_io

InfluxDB data partitioningIsn't a series for every host+service combo excessive?

Because of the way InfluxDB's storage engines

work, no!

We can include the series name as the first part of our

query:

By doing that, InfluxDB can ignore all of the other data in the other series and will only have to access 1

LevelDB file on disk per Grafana query(show Riemann config that creates a series per host+service)

InfluxDB data partitioningBy doing this, we’re not

really querying by hostname or metric name, ie. querying

by those columns

It’s kind of a hack, but it takes advantage of the characteristics of how

InfluxDB’s storage engines work

InfluxDB data partitioningI did this originally to get

around a quirk of Grafana’s UI

For InfluxDB data sources, Grafana doesn’t let you use where or do selects by

more than 1 column

Without splitting data up into more than 1 series, there’s no way to get metric values for an individual host, metric

or host+metric combo

Riemann and InfluxDB data partitioningRiemann’s InfluxDB output function looks like this:

Riemann and InfluxDB data partitioning:series "host.service"

…tells Riemann to take this:{ :host riemann1.local, :service cpu-0.cpu-wait, :metric 3.399911, :tags collectd, :time 1405715017, :ttl 30 }

…and write it to InfluxDB with riemann1.local.cpu-0.cpu-wait as the automatically generated series name

InfluxDB behaves like Graphite with new metrics: it will automatically create a new series if it’s for a hostname.metric

combo it doesn’t already have a series for

GrafanaNow, we just need a dashboard...

Grafana data sourcesBased on Kibana 3 (it's just HTML, JS and CSS)

Deploy it by unzipping a tarball to a place where your webserver can serve the contents

Edit config.js to add data sources:

Grafana can graph data from Graphite, InfluxDB and

OpenTSDB

A new Graphite stack

Riemann

collectdcollectdcollectdcollectd

statsd statsd

InfluxDB

Grafana

Show me some graphs!

(do a demo)

Demo time!

(do an install of Riemann)

(do an install of InfluxDB)

(do an install of Grafana)

Disadvantages of this new stackGrafana doesn't take advantage of all of InfluxDB's features

InfluxDB doesn't have as many built-in functions as graphite-web

Riemann's documentation is sparse, and if you've never written Clojure, there's a learning curve for

writing configs to do more than basic stuff

There are a bunch of dashboard tools that can talk to graphite-web's API

Not as many can talk to InfluxDB

Each of the components are easier to deploy than Graphite and are still being

developed and maintained!

Why use this new stack, then?

InfluxDB and GraphiteInfluxDB actually has a Graphite listener built in

So why use Riemann in the middle? Alerting on metrics!

Riemann actually holds events that it processes in memory for a short time in what it calls the index

Having events in the index means we can keep track of metrics over short periods of time

Riemann and alertingHaving events in the index means we can keep track of

metrics over short periods of time

In the collectd metrics for load average across every machine, calculate an average over the last 5 mins and send an email

alert if it's over a certain threshold

In the collectd metrics for load average on an individual machine, calculate a derivative over the last 5 mins and send

a Nagios alert if the derivative is above a certain level

Native Riemann outputs

Some tools, like CollectD, can output data to Riemann in Riemann's native binary protobuf format

Graphite's format for metrics has become the most commonly used format

For things that only output Graphite format metrics, Riemann's Graphite server functionality is incredibly

useful

Next steps: Scaling Riemann

Because of the way it works internally, Riemann doesn't have support for clustering the way InfluxDB does

https://github.com/jdmaturen/reimann/blob/master/riemann.config.guide#L234

You can set up multiple Riemann servers, and they can forward events to a central one, or one of several behind

HAproxy

Specifically, there's no way to share the in-memory index across nodes

Next steps: Scaling RiemannThis sounds like how you would scale Graphite, but because

we have InfluxDB, each Riemann instance can write to the same InfluxDB instance, or one of many nodes in a cluster

Because InfluxDB can take data in over a network connection, can cluster and is not plain-file-based like Whisper, multiple

Riemanns writing to 1 or more InfluxDBs in a cluster shouldn't be an issue

Some of the monitoring uses of Riemann (calculating moving averages or derivatives) will break because the in-memory

indexes can’t be shared, though

Next steps: Scaling InfluxDBInfluxDB has OotB support for clustering

Uses Raft for a consensus protocol

Sharding is done by blocks of time (time periods are configurable

Metadata is shared via Raft that lets each node know which shards covering what time periods and series/DBs are on

each node

http://sssslide.com/speakerdeck.com/pauldix/the-internals-of-influxdb

Also has a WAL (write ahead log)

Databases and series can be broken up into shards

Tool integrationsRiemann and Logstash can output events to each other

Both can output events to Graphite

This can get really confusing, really quickly

Pies in the skyWhat I want to experiment with and get working:

Email alerts from Riemann

Riemann telling Nagios/Icinga to send alerts based on thresholds for averages or derivatives of metric values

Send web logs to Logstash and make Logstash output metrics from them to Riemann (how many HTTP 404, 503, etc. responses is my web server sending out?)

Linkshttp://riemann.io/

http://influxdb.com/

http://grafana.org/

LinksMonitorama PDX 2014 Grafana workshop:

http://vimeo.com/95316672

Monitorama PDX 2014 InfluxDB talk: http://vimeo.com/95311877

Monitorama Boston 2013 Riemann talk: http://vimeo.com/67181466