Using ClickHouse for Experimentation
13 likes13,321 views
This document discusses using ClickHouse for experimentation and metrics at Spotify. It describes how Spotify built an experimentation platform using ClickHouse to provide teams interactive queries on granular metrics data with low latency. Key aspects include ingesting data from Google Cloud Storage to ClickHouse daily, defining metrics through a centralized catalog, and visualizing metrics and running queries using Superset connected to ClickHouse. The platform aims to reduce load on notebooks and BigQuery by serving common queries directly from ClickHouse.
![170M Monthly Active Users
75M Subscribers
35M Tracks
65 Markets
[1] https://investors.spotify.com
Quick Facts](https://image.slidesharecdn.com/usingclickhouseforexperimentation-180704183700/85/Using-ClickHouse-for-Experimentation-2-320.jpg)
Ad
More Related Content
What's hot (20)
Similar to Using ClickHouse for Experimentation (20)
![[Webinar] Getting Started with BigQuery: Basics, Its Appilcations & Use Cases](https://cdn.slidesharecdn.com/ss_thumbnails/webinargettingstartedwithbigquery-171116122740-thumbnail.jpg?width=560&fit=bounds)
Ad
Recently uploaded (20)
![Download Wondershare Filmora Crack [2025] With Latest](https://cdn.slidesharecdn.com/ss_thumbnails/neo4j-howkgsareshapingthefutureofgenerativeaiatawssummitlondonapril2024-240426125209-2d9db05d-250419-250428115407-a04afffa-thumbnail.jpg?width=560&fit=bounds)
![Get & Download Wondershare Filmora Crack Latest [2025]](https://cdn.slidesharecdn.com/ss_thumbnails/revolutionizingresidentialwi-fi-250422112639-60fb726f-250429170801-59e1b240-thumbnail.jpg?width=560&fit=bounds)
![Pixologic ZBrush Crack Plus Activation Key [Latest 2025] New Version](https://cdn.slidesharecdn.com/ss_thumbnails/fashionevolution2-250322112409-f76abaa7-250428124909-b51264ff-250504160528-fc2bb1c5-thumbnail.jpg?width=560&fit=bounds)
Ad
Using ClickHouse for Experimentation
- 2. 170M Monthly Active Users 75M Subscribers 35M Tracks 65 Markets [1] https://investors.spotify.com Quick Facts
- 3. Organization 1 Company 10 Organizations 30+ Tribes 150+ Squads
- 4. Organization 1 Company 10 Organizations 30+ Tribes 150+ Squads Move fast, break things
- 5. Organization 1 Company 10 Organizations 30+ Tribes 150+ Squads Move fast, break things Ask for forgiveness, not for permission
- 6. Organization 1 Company 10 Organizations 30+ Tribes 150+ Squads Move fast, break things Ask for forgiveness, not for permission AUTONOMY
- 8. ● On-Premise ● 2,500 nodes ● 100 PB Disk ● 100 TB RAM ● 100B+ events per day ● 20K+ jobs per day Hadoop@Spotify
- 9. ● Migration from On-Premise to GCP ● Moved 100 PB of data ● Our Hadoop cluster is dead Hadoop@Spotify
- 12. Randomized Controlled Experiment All Khan Academy content is available for free at www.khanacademy.org
- 13. Randomized Controlled Experiment All Khan Academy content is available for free at www.khanacademy.org
- 14. Randomized Controlled Experiment All Khan Academy content is available for free at www.khanacademy.org
- 15. Randomized Controlled Experiment An experiment where all subjects involved in the experiment are treated the same except for one deviation. One variable is changed in order to isolate the results. All Khan Academy content is available for free at www.khanacademy.org
- 16. A/B Testing A/B Testing is a randomized controlled experiment where one variable is tested. E.g., hypothesis Our new recommendation algorithm increases content consumption. How to verify? 1. Formulate hypothesis 2. Run A/B test 3. See if there is a statistically significant increase in consumption.
- 17. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 Product Owners Data Scientists Granular Data BigQuery
- 18. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 1. Event Delivery Developers instrument applications and services using SDK. Events are collected and published to Pub/Sub. Batch jobs read data from Pub/Sub, deduplicate and anonymize, and then store in hourly partitions on GCS. Exposing users to experiments, and configuring A/B variations on clients is done by dedicates services. Product Owners Data Scientists Granular Data BigQuery 1
- 19. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 2. Data Pipelines and Storage Data gets transformed and aggregated using Dataflow batch jobs, and stored in Bigtable, GCS and BigQuery. Bigtable contains pre-computed aggregated experiment results. BigQuery has granular data used in ad-hoc analysis. Product Owners Data Scientists Granular Data BigQuery 2
- 20. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine 3. Presentation Users of Experimentation platform see their experiment results in web application. Statistical tests and health checks are performed automatically. Metrics for Experimentation Platform v1 Product Owners Data Scientists Granular Data BigQuery 3
- 21. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 4. Ad-hoc Analytics Data scientists do ad-hoc exploration in Jupyter notebooks using BigQuery. Here they answer experiment specific-questions, not automatically supported by experimentation system. Product Owners Data Scientists Granular Data BigQuery 4
- 22. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 What works well Centralized team owning 100-s of core metrics. Automatic experiment analysis and planning. Allows to conclude experiments without manual analysis. Autonomous feature teams can move fast and iterate on their product. Product Owners Data Scientists Granular Data BigQuery
- 23. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Aggregated Data Cloud Bigtable Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 Problems Not every metric worths centralization. Centralized team became a bottleneck for Feature features. As a result, too much repetitive work goes into notebooks and ad-hoc queries. Product Owners Data Scientists Granular Data BigQuery
- 24. Product TeamsEvent Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Storage Granular Data OLAP Database Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Compute Engine Metrics for Experimentation Platform v1 Reasons 1. Experimentation isn’t only about hypothesis testing, but learning from experiments. Aggregated data in Bigtable wasn’t granular enough, and didn't have enough dimensions. 2. Can’t add a new metric without involving a central team. What we want Provide teams more granular data out of the box, and give a way to define a new metric. Product Owners Data Scientists Granular Data BigQuery
- 25. Requirements ● Serve 100-s of QPS with sub-second latency ● We know in advance what are queries and data ● Maintain 10x metrics with the same cost ● Thousands of metrics ● Billions of rows per day in each of 100-s of tables ● Ready to be used out of the box ● Leverage existing infrastructure as much as feasible ● Hide unnecessary complexity from internal users
- 26. What about BigQuery? ● Supports Standard SQL ● Don’t have to optimize datasets in advance ● Works great for heavy queries with joins among multiple datasets ● Doesn’t need operations and machines running ● Good for interactive ad-hoc queries (~ minutes) ● Isn’t best for a high amount of low-latency queries you are aware in advance
- 27. Why ClickHouse? ● Build proof of concept using various OLAP storages (ClickHouse, Druid, Pinot, ...) ● ClickHouse has the most simple architecture ● Powerful SQL dialect close to Standard SQL ● A comprehensive set of built-in functions and aggregators ● Was ready to be used out of the box ● Superset integration is great ● Easy to query using clickhouse-jdbc and jooq
- 28. Event Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Product TeamsStorage Granular Data ClickHouse Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Metrics for Experimentation Platform v2 5. ClickHouse Interactive queries on granular data. Reduce demand in notebooks and BigQuery with dashboards and exploration in Superset. Product Owners Data Scientists Granular Data BigQuery Superset 5
- 29. Event Delivery Cloud Pub/Sub Cloud Storage Cloud Dataproc Data Pipelines Cloud Dataflow Product TeamsStorage Granular Data ClickHouse Granular Data Cloud Storage Ad-hoc Analytics Presentation Web Application Metrics for Experimentation Platform v2 6. Metrics Catalog Centralized place for teams to define their own metrics. 20 minutes to define a metric. Product Owners Data Scientists Granular Data BigQuery Superset Metrics Catalog Metrics API Metric definitions 6
- 30. What we have built ● Own DSL to define metrics, and centralized metrics catalog ● Expressive and simple model that we can efficiently scale to 1000-s of metrics ● Generalize existing components to work with Metrics DSL ○ data preparation and ingestion into ClickHouse ○ denormalization with conformed dimensions ○ create dashboards, tables and charts in Superset ○ do statistical tests, and expose results through API ○ define ownership, tiering, and other attributes ○ integrates with the rest of infrastructure for alerting, monitoring, data quality, anomaly detection, access control & etc ● Users don’t work with ClickHouse SQL, or need to know how it works ● API to query metrics and metadata
- 31. Ingestion to ClickHouse ● Move data from GCS to ClickHouse ● Use clickhouse-jdbc, custom code and RowBinary format ● Use daily partitioning, and ingest once a day ● 1 hour to ingest 5 TiB on test cluster using 9 n1-standard-32 with 8 NVMe SSD RAID0 ● Don’t use materialized views in ClickHouse ● Offload most of computations to batch data pipelines due to scalability, experience and tooling ● TODO try ClickHouse-Native-JDBC ● TODO pre-sort in data pipelines before ingesting
- 32. What is next ● Do lambda-style ingestion for subset of metrics with low-latency requirements ● Add more aggregations to DSL (e.g. 5 statistical moments) ● Add custom chart types to Superset ● Try ClickHouse for similar use cases within Spotify