This write-up is a deep insight into the persistence layer of Twitter. It does answer all our questions about it in-depth such as What database does it use? Is it SQL, NoSQL, a polyglot persistent system? How does it store hundreds of millions of tweets sent every single day?
I’ll begin with the short answer first, list out the databases & then will talk about each one of them in detail.
So, without further ado.
Let’s get on with it.
1. What Database Does Twitter Use?
Hadoop for social graph analysis, recommendations, trends, API analytics, user engagement prediction, Ad targeting, ad analytics, Tweet impressions processing, taking MySQL backups, Manhattan backups & storing front end scribe logs.
MySQL & Manhattan as the primary data stores for storing user data.
Memcache, Redis for caching
FlockDB for storing social graph
MetricsDB for storing platform data metrics
Blobstore for storing images, videos & large binary objects.
Now let’s talk about all the data stores individually.
Twitter runs one of the biggest Hadoop clusters in the world. Hadoop was originally used at Twitter to take MySQL backups & then over time its use cases on the platform grew by quite an extent.
Today, It is used to store data to run analytics on the actions the users perform on the platform. Social graph analysis, recommendations, trends, API analytics, user engagement prediction, Ad targeting, ad analytics, Tweet impressions processing, taking MySQL backups, Manhattan backups & storing Front end scribe logs.
The Hadoop file system stores over 500 Petabytes of data running over tens of thousands of instances. The entire cluster is managed using the Hadoop Federation feature running hundreds of thousands of daily Hadoop jobs. Tens of millions of daily Hadoop tasks.
150K+ services & 130M containers are running on any regular day.
To scale the services Hadoop uses different namespaces. The nodes with namespaces, also known as Namenodes, run independently & have no dependence on each other.
There are nodes called data nodes which are used for data storage by the Namenodes. These data nodes are registered with the Namenodes in the cluster.
To manage the namespaces, View file system is used, which is useful for managing clusters having multiple Namenodes.
Every cluster has more than 3500 Namenodes.
3. Manhattan NoSQL Eventually Consistent Data Store
Manhattan is a real-time multi-tenant scalable distributed database used at Twitter to serve millions of queries per second with really low latency being highly available.
It stores user tweets, direct messages, account details & much more. On any given day, Manhattan clusters handle 10s of millions of queries per second.
The real-time nature of the platform requires very low latency which the existing open-source products couldn’t offer.
A lot of resources were spent on managing & trying to stoke up the performance of the existing storage systems. This prognosis wasn’t good, the existing persistence system wouldn’t have worked in the long run. Which led to the development of Manhattan.
The requirements to build Manhattan were simple like having a reliable datastore with predictable performance, availability, extensibility, easy operability working with hundreds of thousands of nodes & scalability.
Manhattan is separated into four layers interfaces, storage services, storage engines and the core with an eventual consistency model.
High availability is preferred over strong consistency. CAP theorem
Eventually consistency model fits best in social networks, it allows the services to scale seamlessly. Primarily all of the NoSQL data stores follow an eventual consistency model which enables them to scale on the fly.
Replication is extensively used to keep Manhattan highly available.
The storage engine of Manhattan was designed in-house with a pluggability feature to enable plugging of any external storage engines if & when required in future.
4. BlobStore – Image Video & Large Object Storage System
Blobstore is Twitter’s scalable storage system to store user images, videos & other large binary objects. It has enabled Twitter to cut down costs associated with storing user uploaded images with Tweets.
It is a highly performant system capable of serving images in low tens of milliseconds while having a throughput of hundreds of thousands of requests per second.
When the image is uploaded to the Blobstore, it synchronizes the image across all of Twitter’s data centres with the help of an asynchronous queue server.
5. Caching at Twitter With Memcache & Redis
With it’s caching Twitter delivers approx. 120 GB of data per second to the clients.
Twitter built Twemcache, a custom version of Memcache suited for large scale production deployment.
The platform has hundreds of cache instances having over 20TB of data in-memory from over 30 services. Comprehensively, the caching layer serves over 2 trillion queries on any regular day.
Caching at Twitter has two primary use cases
The first is the obvious one, having in-memory data to avoid database hits. The second, it is used as a memory buffer to store items that are expensive to compute over and over.
Twemproxy, a lightweight proxy for Memcache & Redis protocol, enables the caching layer to scale horizontally minimizing the connections to the backend caching servers.
Twitter internally runs many caching services & one of them is powered by Redis. Redis clusters cache direct user messages, ad spends, their impressions & engagement.
There is another caching service, Haplo which acts as a primary cache for the Twitter Timeline. It’s backed by Redis.
To educate yourself on software architecture from the right resources, to master the art of designing large scale distributed systems that would scale to millions of users, to understand what tech companies are really looking for in a candidate during their system design interviews. Read my blog post on master system design for your interviews or web startup.
6. Relational Databases – MySQL, PostgreSQL, Vertica
Twitter started with MySQL as the primary data store, from a single instance the persistence layer grew to a large number of clusters.
Twitter has one of the biggest deployments of MySQL right from its inception. It has MySQL clusters with thousands of nodes serving millions of queries per second.
MySQL has primarily two use cases
1. Acting as the storage node for the distributed data store within Twitter’s sharding framework. MySQL storage nodes provide reliability & performance in the overall distributed store.
2. Powers services such as ads, authentication, twitter trends & many internal services.
The engineering team at Twitter has also built a scalable MySQL database service framework called Mysos.
It is based on Apache Mesos. Mesos enables Mysos to schedule, monitor & communicate with MySQL instances in the cluster.
MySQL instances were sharded with the help of a framework for creating distributed datastores called Gizzard.
Besides MySQL other SQL based data stores like PostgreSQL & Vertica are also used to store ad campaigns, sales & internal tooling data.
Cassandra was also used with MySQL as a data storage solution but it lacked the auto-increment feature which MySQL had. It did store metrics for a while & was later deprecated after Twitter started using Manhattan.
7. Metrics DB – Time Series Database For Storing Metrics
Metrics DB is used at Twitter to store the metrics. The metric ingestion rate is more than 5 billion metrics per minute with 25K query requests per minute.
Originally, Manhattan was used as the metric storage database but Twitter faced scalability issues along with not having support for additional minute metric tags.
This led to the development of Metrics DB in-house using the compression algorithm of Facebook’s in memory Time-series Database Gorilla.
Metrics DB provides multi-zone support, partitioning of metrics & compression efficiency over other data stores used at Twitter. Using the Gorilla’s compression algorithm Twitter reduced its space usage by 95%.
8. Social Graph Storage with FlockDB
FlockDB is a distributed graph datastore built for fast graph traversals, storing adjacency lists, supporting a high rate of add remove update operations, paginating through millions of entries, horizontally scaling, running graph walking queries.
Twitter used it to store, social graph, information such as who follows whom & stuff. But the GitHub repository says that Twitter is no longer maintaining this project. I am not sure if they still use it to store their social graph.
Recommended Read: Master System Design For Your Interviews Or Your Web Startup
Subscribe to the newsletter to stay notified of the new posts.
Well, Guys!! This is pretty much it about the architectural design & the technology stack of Instagram. If you liked the write-up, share it with your folks. Consider following 8bitmen on Twitter, Facebook, LinkedIn to stay notified of the new content published.
I am Shivang, the author of this writeup. You can read more about me here.
9. More On the Blog
> Spotify Engineering: From Live to Recording
> Ingesting LIVE video streams at a global scale at Twitch
> $64,944 spent on AWS, to support 25,000 customers, in August by ConvertKit.
> Read how Storytel engineering computes customer consumption of books transitioning from batch processing to streaming bookmarks data with Apache Beam and Google Cloud.
> How Pokemon Go scales to millions of requests per second?
> Insight into how Grab built a high-performance ad server.
SUBSCRIBE TO MY NEWSLETTER to be notified of new additions to the list. Fortnight/monthly emails.
Looking for developer, software architect jobs? Try Jooble. Jooble is a job search engine created for a single purpose: To help you find the job of your dreams!!
- Web Application Architecture Explained With Designing a Real-World Service
- Wide-column, Column-oriented and Column Family Databases – A Deep Dive with Bigtable and Cassandra
- Design For Scale and High Availability – What Does 100 Million Users On A Google Service Mean?
- How Razorpay handled significant transaction bursts during events like IPL
- Facebook’s Photo Storage Architecture