Rounds analytics pipeline
5 likesâ˘1,870 views
Rounds Analytics, founded in 2008 and based in Tel Aviv, has raised over $22 million and serves over 25 million users. The document outlines their transition from a monolithic app to a microservices architecture for user analytics, leveraging tools like Google BigQuery, Elasticsearch, and a variety of programming languages. It highlights the challenges faced in data collection and ETL processes, and the current setup for real-time analytics and monitoring.
Rounds analytics pipeline
- 1. Rounds Analytics Pipeline Aviv Laufer CTO, Rounds @avivl
- 2. Founded 2008 35 person team (â in R&D) Tel-Aviv based Raised over $22 million in funding from industry leading investors such as Sequoia Capital, Samsung Ventures, Rhodium, Verizon Ventures and many more. Over 25 million users worldwide
- 4. Client - Objective-C, Java, C/C++ Server - Python, Go, C++, and bits of Erlang DB - MySQL (RDS), CouchBase (a few clusters) Multi Cloud - AWS, GCE, SL, DO Deployment - Ansible Monitoring - Sensu, NewRelic, VictorOps And yes we use DockerâŚ... Tools of the Trade
- 5. We tried to make it work quite a few times, and failed We kept on trying We think we got it right this time ANALYTICS @ROUNDS DO OR DO NOT. THERE IS NO TRY. - Master Yoda
- 6. One monolith app Data was written to MySql/RDS - row by row Batch ETL to Vertica And then came July 2014 GENESIS
- 8. Data collection killed our backend app Slow, failing ETL process No real time view into events Preferred users over analytics, we killed the event collection We were flying blind CHAOS
- 9. Separate ETL process from main app Clients reports (a request for each event) to a different microservice First very naive version written in Go - it scales! Data is written to an Elasticsearch cluster. ETL from ES to Vertica EXODUS
- 10. Frontends - Receiving user analytics and perform sanity checks Google Pub/Sub - Store events for future processing Workers - Pull the events from Pub/Sub and stream to Google BigQuery and ES ...And Then There Were Three...
- 11. Clients send gzipped, batched Frontend does sanity checks - Validation, versioning, etc. Frontend replies fast (202 Accepted) and closes the connection in order to save on mobile socket life Geo load-balanced Pushes analytics into Pub/Sub for future processing, mutation Fan-In model ANALYTICS - FRONTEND
- 12. Pulls analytics from Pub/Sub Mutate/Enrich data if necessary Inserts to various DBs, according to usage - Monitoring, BI, Warehousing, etc. Separation of concerns - Worker cluster per target DB Fan-out model Renee Finch, golang.org/doc/gopher/pencil/ANALYTICS - WORKER
- 13. Golang, abstraction package Receives rows, streams to BigQuery (as opposed to load jobs) Sync (foreground insert) or Async (background insert) Pros: Instant data availability, no job delay, fast Cons: Harder handling of bad analytics, Googleâs HTTP 500s (requires retry) Open source, PRs merrily encouraged! Collecting User Data and Usage - Blog Post - http://bit.ly/CollectingDataRounds github.com/rounds/go-bqstreamerSTREAMING TO BigQuery
- 14. Frontends deployed in several locations in GCE (We Geo load balance them) Workers are in GCE (Europe West) ES cluster is in GCE (Europe West) ACROSS THE UNIVERSE
- 15. We started using Elasticsearch for monitoring about a year before elastic.co relaize that Every new feature received a monitoring dashboard Debugging Monitoring (custom sensu checks) Data is kept for 30 to 90 days Ad-hoc reporting using Kibana ELASTICSEARCH
- 17. Store data from the beginning till the end of time Standard(ish) SQL Very fast No DBA is needed Business reports (SiSense) - âKibanaâ for BigQuery BigQuery
- 18. NEW ISSUES Permissive vs hard scheme for events - allow clients ease of use while keeping the scheme strict for ease of BI Clients make mistakes (Arabic locale dates) - elasticsearch allows while BQ doesnât Things weâre integrating as solutions Every event is a class - compile time validation generated from RAML Wrote a library for event reporting server side
- 19. ANY QUESTIONS DO YOU HAVE?