https://www.youtube.com/watch?v=VBIlk3GMmNc&list=PL_arKaS8JOXHgSDJXtAVxhtDBJ6WL3uaS&index=16
AppsFlyer’s mobile attribution and analysis platform is used by the biggest and most popular applications on Earth, generating a constant “storm” of 70B+ events (HTTP Requests) on their microservices, cloud based platform daily. In this talk, Morri will share their technological choices which include Clojure as their leading backend language - and the decisions to migrate from Python for improved multi-threading and concurrency.
The backend was to built to be a robust system based on a diversity of open source tooling such as: Kafka, RabbitMQ, Aerospike, Redis and a host of proprietary in-house developed tools and services that enable the testing and adoption of new data technologies, continuous deployment, and large-scale monitoring of the system - including open sourcing production libraries for interoperability with core technologies.
This talk will also dive into AppFlyer’s real-time back-end architecture & functional programming philosophy, what it is like to be a developer at AppsFlyer, and overall attitude towards performance, redundancy and resiliency for processing 35 Million events/minute at an average latency of hundreds of milliseconds per event.
2 -person team
- Python
- Redis
- MongoDb (
- CouchDB (master database)
over time
- Clojure (primary language) principles of language permeated to architecture; Language of the system Rich
- Kafka
- Docker, Consul, Spark
- Google BigQuery
- Aerospike
- Druid
- Go
- ClickHouse
- Kafa, aerospike, clickhouse
Mainly a data platform with few microservices (attribution, user, postbacks, push) than a SOA Mobile phones hitting web servers. webservers puts into kafka
- Microservices pulls from kafka
- Attribution Engine + Aerospike
- User Activity + Aerospike
- Postbacks to Partners
- Push API to Clients
- Batch Analytics + Druid + ClickHouse
- Also S3 for retention beyond kafka
- Event Driven Architecture and also a mini-CQRS (both have Kafka as central DB)
- JSON/Protobuf on wire. Parquet for S3. Parquet is column oriented. so gives rigidity
- No rename of data fields is the
- downstream people pains (data teams) cannot be heard (app team)
- DB is an implementation detail
- spot instances
- Kafka - 280 nodes, 15 clusters
- Aerospike - 220 nodes, 17 clusters
- Apache Spark - 2600 nodes, 18 clusters
- Amazon S3 - 12 PB
- Release It! - Michael T Nygard