An Introduction to Prometheus (GrafanaCon 2016)

Introduction to Prometheus
An Approach to Whitebox Monitoring

Who am I?
Engineer passionate about running software reliably in production.
Studied Computer Science in Trinity College Dublin.
Google SRE for 7 years, working on high-scale reliable systems.
Contributor to many open source projects, including Prometheus, Ansible,
Python, Aurora and Zookeeper.
Founder of Robust Perception, provider of commercial support and consulting
for Prometheus.

Blackbox monitoring
Monitoring from the outside
No knowledge of how the application works internally
Examples: ping, HTTP request, inserting data and waiting for it to appear on
dashboard

Where to use Blackbox
Blackbox monitoring should be treated similarly to smoke tests.
It’s good for finding when things have badly broken in an obvious way, and testing
from outside your network.
Not so good for knowing what’s going on inside a system.
Nor should it be treated like regression testing and try to test every single feature.
Tend to be flaky, as they either pass or fail.

Whitebox Monitoring
Complementary to blackbox monitoring.
Works with information from inside your systems.
Can be simple things like CPU usage, down to the number of requests triggering a
particular obscure codepath.

Prometheus
Inspired by Google’s Borgmon monitoring system.
Started in 2012 by ex-Googlers working in Soundcloud as an open source project.
Mainly written in Go. Version 1.0 released in 2016. Incubating with the CNCF.
500+ companies using it including Digital Ocean, Ericsson, Weave and CoreOS.

Why monitor?
Know when things go wrong
Be able to debug and gain insight
Trending to see changes over time
Plumbing data to other systems/processes

Knowing when things go wrong
The first thing people think of you say monitoring is alerting.
What is the wrongness we want to detect and alert on?
A blip with no real consequence, or a latency issue affecting users?

Symptoms vs Causes
Humans are limited in what they can handle.
If you alert on every single thing that might be a problem, you'll get overwhelmed
and suffer from alert fatigue.
Key problem: You care about things like user facing latency. There are hundreds
of things that could cause that.
Alerting on every possible cause is a Sisyphean task, but alerting on the symptom
of high latency is just one alert.

Example: CPU usage
Some monitoring systems don't allow you to alert on the latency of your servers.
The closest you can get is CPU usage.
False positives due to e.g. logrotate running too long.
False negatives due to deadlocks.
End result: Spammy alerts which operators learn to ignore, missing real problems.

Many Approaches have Limited Visibility

Monitor as a Service, not as Machines

Freedom for Alerting
A system like Prometheus gives you the freedom to alert on whatever you like.
Alerting on error ratio across all the machines in a datacenter? No problem.
Alerting on 95th percentile latency for the service being <200ms? No problem.
Alerting on data taking too long to get through your pipeline? No problem.
Alerting on your VIP not giving the right HTTP response codes? No problem.
Produce alerts that require intelligent human action!

Debugging to Gain Insight
After you receive an alert notification you need to investigate it.
How do you work from a high level symptom alert such as increased latency?
You drill down through your stack with dashboards to find the subsystem that's the
cause!

Metrics from All Levels of the Stack
Many existing integrations: Java, JMX, Python, Go, Ruby, .Net, Machine,
Cloudwatch, EC2, MySQL, PostgreSQL, Haskell, Bash, Node.js, SNMP, Consul,
HAProxy, Mesos, Bind, CouchDB, Django, Mtail, Heka, Memcached, RabbitMQ,
Redis, RethinkDB, Rsyslog, Meteor.js, Minecraft and Factorio.
Graphite, Statsd, Collectd, Scollector, Munin, Nagios integrations aid transition.
It’s so easy, most of the above were written without the core team even knowing
about them!

Metrics are just one Tool
Metrics are good for alerting on issues and letting you drill down the focus of your
debugging.
Not a panacea though, as with all approaches fundamental limitations on data
volumes.
For successful debugging of complex problems you need a mix of logs, profiling
and source code analysis.

Trending and Reporting
Alerting and debugging is short term.
Trending is medium to long term.
How is cache hit rate changing over time?
Is anyone still using that obscure feature?
With Prometheus you can do analysis beyond this.

Powerful Query Language
Can multiply, add, aggregate, join, predict, take quantiles across many metrics in
the same query. Can evaluate right now, and graph back in time.
Answer questions like:
What’s the 95th percentile latency in each datacenter over the past month?
How full will the disks be in 4 days?
Which services are the top 5 users of CPU?

Example: Top 5 Docker images by CPU
topk(5,
sum by (image)(
rate(container_cpu_usage_seconds_total{
id=~"/system.slice/docker.*"}[5m]
)
)
)

Structured Data: Labels
Prometheus doesn’t use dotted.strings like metric.grafnacon.nyc.
Multi-dimensional labels instead like
metric{event=”grafanacon”,aircraft_carrier_location=”nyc”}
Can aggregate, cut, and slice along them.
Can come from instrumentation, or be added based on the service you are
monitoring.

Example: Labels from Node Exporter

Python Instrumentation: An example
pip install prometheus_client
from prometheus_client import Summary, start_http_server
REQUEST_DURATION = Summary('request_duration_seconds',
'Request duration in seconds')
@REQUEST_DURATION.time()
def my_handler(request):
pass // Your code here
start_http_server(8000)

Adding Dimensions (No Evil Twins Please)
from prometheus_client import Counter
REQUESTS = Counter('requests_total',
'Total requests', ['method'])
def my_handler(request):
REQUESTS.labels(request.method).inc()
pass // Your code here

Labels go beyond Prometheus
If you're using Kubernetes, Prometheus can take in your labels and annotations
too.
Similar data models and mutual integrations make your life easier!

Plumbing
Prometheus isn't just open source, it's also an open ecosystem.
We know we can't support everything, so at every level there's a generic interface
to let you get data in and/or out.
So for example if you want to run a shell script when an alert fires, you can make
that happen.

Prometheus Clients as a Clearinghouse

Monitoring What Matters with Prometheus
To summarise, the key things Prometheus empowers you to build:
Alerting on symptoms. Alerts which require intelligent human action.
Debugging dashboards that let you drill down to where the problem is.
The ability to run complex queries to slice and dice your data.
Easy integration points for other systems.
These are good things to have no matter which monitoring system(s) you use.

10 Tips for Monitoring
With potentially millions of time series across your system, can be difficult to know
what is and isn't useful.
What approaches help manage this complexity?
How do you avoid getting caught out?
Here's some tips.

#1: Choose your key statistics
Users don't care that one of your machines is short of CPU.
Users care if the service is slow or throwing errors.
For your primary dashboards focus on high-level metrics that directly impact
users.

#2: Use aggregations
Think about services, not machines.
Once you have more than a handful of machines, you should treat them as an
amorphous blob.
Looking at the key statistics is easier for 10 services than 10 services each of
which is on 10 machines
Once you have isolated a problem to one service, then can see if one machine is
the problem

#3: Avoid the Wall of Graphs
Dashboards tend to grow without bound. Worst I've seen was 600 graphs.
It might look impressive, but humans can't deal with that much data at once.
(and they take forever to load)
Your services will have a rough tree structure, have a dashboard per service and
talk the tree from the top when you have a problem. Similarly for each service,
have dashboards per subsystem.
Rule of Thumb: Limit of 5 graphs per dashboard, and 5 lines per graph.

#4: Client-side quantiles aren't aggregatable
Many instrumentation systems calculate quantiles/percentiles inside each
process, and export it to the TSDB.
It is not statistically possible to aggregate these.
If you want meaningful quantiles, you should track histogram buckets in each
process, aggregate those in your monitoring system and then calculate the
quantile.
This is done using histogram_quantile() and rate() in Prometheus.

#5: Averages are easy to reason about
Q: Say you have a service with two backends. If 95th percentile latency goes up
due to one of the backends, what will you see in 95th percentile latency for that
backend?
A: ?

#5: Averages are easy to reason about
Q: Say you have a service with two backends. If 95th percentile latency goes up
due to one of the backends, what will you see in 95th percentile latency for that
backend?
A: It depends, could be no change. If the latencies are strongly correlated for each
request across the backends, you'll see the same latency bump.
This is tricky to reason about, especially in an emergency.
Averages don't have this problem, as they include all requests.

#6: Costs and Benefits
1s resolution monitoring of all metrics would be handy for debugging.
But is it ten time more valuable than 10s monitoring? And sixty times more
valuable than 60s monitoring?
Monitoring isn't free. It costs resources to run, and resources in the services being
monitored too. Quantiles and histograms can get expensive fast.
60s resolution is generally a good balance. Reserve 1s granularity or a literal
handful of key metrics.

#7: Nyquist-Shannon Sampling Theorem
To reconstruct a signal you need a resolution that's at least double it's frequency.
If you've got a 10s resolution time series, you can't reconstruct patterns that are
less than 20s long.
Higher frequency patterns can cause effects like aliasing, and mislead you.
If you suspect that there's something more to the data, try a higher resolution
temporarily or start profiling.

#8: Correlation is not Causation - Confirmation Bias
Humans are great at spotting patterns. Not all of them are actually there.
Always try to look for evidence that'd falsify your hypothesis.
If two metrics seem to correlate on a graph that doesn't mean that they're related.
They could be independent tasks running on the same schedule.
Or if you zoom out there plenty of times when one spikes but not the other.
Or one could be causing a slight increase in resource contention, pushing the
other over the edge.

#9 Know when to use Logs and Metrics
You want a metrics time series system for your primary monitoring.
Logs have information about every event. This limits the number of fields (<100),
but you have unlimited cardinality.
Metrics aggregate across events, but you can have many metrics (>10000) with
limited cardinality.
Metrics help you determine where in the system the problem is. From there, logs
can help you pinpoint which requests are tickling the problem.

#10 Have a way to deal with non-critical alerts
Most alerts don't justify waking up someone at night, but someone needs to look
at them sometime.
Often they're sent to a mailing list, where everyone promptly filters them away.
Better to have some form of ticketing system that'll assign a single owner for each
alert.
A daily email with all firing alerts that the oncall has to process can also work.

Questions?
Project Website: prometheus.io
Demo: demo.robustperception.io
Company Website: www.robustperception.io

An Introduction to Prometheus (GrafanaCon 2016)

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to An Introduction to Prometheus (GrafanaCon 2016) (20)

More from Brian Brazil (9)

Recently uploaded (19)

An Introduction to Prometheus (GrafanaCon 2016)