Deploying prometheus is easy and running single instance can be sufficient for most deployments. We will talk about scalability limits of prometheus instance, when and how use shardIng, what is trickster and why you should use it, too and how thanos can help you when all hope is lost.

1. Monitoring with Prometheus
at Scale
Adam Hamsik
adam.hamsik@lablabs.io

2. Labyrinth Labs
Rock-solid infrastructure and DevOps
● Building rock-solid and secure foundations for all your digital operations. Our
mission is to let you focus on your business without ever needing to worry
about technical issues again.
● Making you ready for growing traffic, safe against new security vulnerabilities
and data-loss.
2

3. TL;DR
● We will start with common monitoring issues and problems.
● Deploying Prometheus is easy and running a single instance can be sufficient
for most deployments.
● We will have a quick look at AlertManager
● We will talk about scalability limits of prometheus instance, when and how to
use sharding.
● What is Trickster and why you should use it too
● How Thanos/Cortex can help you when all hope is lost.
3

4. Common Monitoring Problems
● Monitoring tools are limited both technically and conceptually
● Most of existing tools don’t really scale with current infrastructure needs.
● Limited visibility
○ Generally we want to monitor and gather as much information as we can.
○ Even if we don’t need it right away usually it will be useful in a future(I promise)
● No common application monitoring interface. There are different
protocols/standards
○ Openmetrics
○ SNMP
4

5. Common Monitoring Problems
5

6. Prometheus Monitoring System
The Prometheus monitoring system and time series database is CNCF graduated
project.
● Originally developed by exGooglers for SoundCloud as their internal monitoring
system
● Inspired by Google’s Borgmon monitoring system
● Open Source under the Apache License
● Written as monolithic application in Go
6

7. Prometheus Server Overview
● Multi-dimensional data model with time series data identified by metric
name and key/value(labels) pairs
● PromQL, a flexible query language to leverage this dimensionality
● No reliance on distributed storage; single server nodes are autonomous
● Targets are discovered via service discovery or static configuration
● Pushing time series is supported via an intermediary gateway
● Monitor Services not Machines/Servers
7

8. Prometheus Architecture
81. https://www.prometheus.io/assets/architecture.png

9. Company Prometheus
Usage
● We deployed first prometheus servers
● Add some services
● Setup trickster as a Grafana Cache
● Add more services/servers
● Continuous adding of CPU/Memory to Prometheus instance
● Setup simple federation/sharding if single instance is too big
● Use Thanos
9

10. First Prometheus Deployment
● Deploying your first Prometheus server is very easy. Fetch prometheus
binary + config.
● There is a no concept of a Prometheus Cluster
● Generally Prometheus can scale very well with CPU/Memory
○ Providing more cpu/memory allows prometheus to monitor more
metrics
○ It’s hard to run large pod in a kubernetes cluster if it’s as big as a
worker node.
● If job is too big for a single server you can use federation/sharding
(remote reads) for simple scaling
10

11. Trickster Setup
● Loading complicated/big dashboard on Grafana can overload your
prometheus server
○ Use trickster to cache PromQL results for future reuse
○ Queries on metrics with high cardinality can use a lot of memory on
you prometheus instance[1].
○ Use limits to make sure user will not overload your server
query.max-concurrency/query.max-samples
● Delta Proxy caching - inspects the time range of a client query to
determine what data points are already cached
111. https://www.robustperception.io/limiting-promql-resource-usage

12. Trickster Setup
121. https://github.com/tricksterproxy/trickster/blob/main/docs/images/partial-cache-hit.png

13. Trickster Setup
131. https://secure.meetupstatic.com/photos/event/5/7/9/e/600_469882430.jpeg

14. Metrics Cardinality
● Prometheus performance almost always comes to one thing metrics
cardinality.
● Cardinality describes how many unique values of some metric you have
○ container_tasks_state metric will have a unique (pod/container) pair for each running
container in your cluster
○ custom_api_http_request will have a unique metric for each combination of
url/http_method/env. (/api/v2/users, get, dev; /api/v2/users, post, prod...)
141. https://www.robustperception.io/cardinality-is-key

15. Bad Metrics Cardinality
151. https://www.robustperception.io/cardinality-is-key
● See example below where we throw away bad fluentd metrics and dropped number of
scrapped metrics by ½
● If you are using fluentd look for fluentd_tail_file_inode, fluentd_tail_file_position
○ In our use case we saw cardinality 1220 from 2 metrics above per node !

16. Thanos/Cortex as ultimate solution
● If you have multiple kubernetes clusters, datacenters with millions of
metrics and adding more CPU/memory to prometheus is not an option.
○ Consider adding Thanos/Cortex to your infrastructure
● Thanos querier Prometheus Server HA, can load metrics from multiple
prometheus servers and make sure it will present full data to user.
○ Implements Prometheus 1.1 HTTP api.
● Thanos compactor can downsample, change retention or resolution of
your metrics.
● Thanos store is a component which can save your metrics in a AWS S3
compatible object store.
16

17. Thanos architecture
17

18. Thanos SideCar
18
● It implements Thanos’ Store API on top of Prometheus’ remote-read API. This allows
Queriers to treat Prometheus servers as yet another source of time series data without
directly talking to its APIs.
● Optionally, the sidecar uploads TSDB blocks to an object storage bucket as Prometheus
produces them every 2 hours. This allows Prometheus servers to be run with relatively
low retention while their historic data is made durable and queryable via object storage.
● Optionally Thanos sidecar is able to watch Prometheus rules and configuration,
decompress and substitute environment variables if needed and ping Prometheus to
reload them.

19. Thanos Query
19
● The PromQL query is posted to the Querier
● It interprets the query and goes to a pre-filter
● The query fans out its request for stores, prometheuses or other queries on the basis of labels and
time-range requirements
● The Query only sends and receives StoreAPI messages
● After it has collected all the responses, it merges and deduplicates them (if enabled)
● It then sends back the series for the user
1. https://banzaicloud.com/img/blog/multi-cluster-monitoring/life_of_a_query.png

20. Questions ?
20

21. Thank You.
We are hiring, remote working DevOps/Kubernetes
engineers.
adam.hamsik@lablabs.io
www.lablabs.io
21