Strengthening resiliency from the scale from the Tinder that have Amazon ElastiCache

It is a visitor post out-of William Youngs, App Engineer, Daniel Alkalai, Older Software Professional, and you will Jun-more youthful Kwak, Elder Systems Movie director with Tinder. Tinder is introduced with the a college campus within the 2012 that will be the latest planet’s most popular app for appointment new people. It’s been installed over 340 million times and that is obtainable in 190 regions and forty+ languages. By Q3 2019, Tinder had almost billion subscribers and is the best grossing low-gambling application around the world.

Within Tinder, i rely on the reduced latency regarding Redis-oriented caching so you’re able to service dos billion everyday representative tips if you are hosting more than 29 million fits. The majority of all of our research businesses is reads; the next diagram illustrates the overall data disperse buildings of our backend microservices to create resiliency during the scale.

Inside cache-out approach, when a microservices receives a request studies, it requests a beneficial Redis cache towards the study earlier drops to a resource-of-knowledge chronic databases shop (Amazon DynamoDB, however, PostgreSQL, MongoDB, and you may Cassandra, are now and again made use of). Our features next backfill the importance to your Redis on resource-of-specifics in case there are a good cache skip.

Before i adopted Craigs list ElastiCache to have Redis, we made use of Redis hosted to the Craigs list EC2 era with app-depending customers. We accompanied sharding from the hashing tips predicated on a static partitioning. The fresh diagram over (Fig. 2) depicts good sharded Redis configuration to your EC2.

Specifically, the software readers managed a predetermined setup of Redis topology (including the quantity of shards, level of reproductions, and like proportions). All of our applications then utilized this new cache data on top of an effective considering repaired configuration outline. The fixed repaired setting required in so it services brought about tall activities towards the shard addition and single muslim seznamka rebalancing. Still, this notice-then followed sharding service performed relatively well for all of us in early stages. not, just like the Tinder’s popularity and request guests increased, very did what number of Redis occasions. So it increased the latest over as well as the challenges out of maintaining her or him.


Earliest, the fresh new working load off keeping all of our sharded Redis class are is tricky. They took a lot of innovation time for you manage our very own Redis groups. Which above put-off extremely important technology operate our engineers have focused on instead. Such as for instance, it absolutely was an enormous experience so you’re able to rebalance clusters. We must copy an entire class merely to rebalance.

Second, inefficiencies inside our implementation requisite infrastructural overprovisioning and you will increased expense. The sharding formula is inefficient and you may contributed to scientific complications with beautiful shards very often called for developer input. Additionally, if we necessary our very own cache study as encoded, we’d to make usage of the encryption our selves.

Eventually, and most importantly, all of our yourself orchestrated failovers caused application-broad outages. The brand new failover away from a great cache node this one your key backend attributes utilized was the cause of connected services to shed the associations to the node. Till the software was restarted in order to reestablish connection to the mandatory Redis such as for instance, all of our backend assistance were often entirely degraded. This is probably the most high motivating foundation in regards to our migration. Just before the migration so you’re able to ElastiCache, the fresh new failover from an excellent Redis cache node try the largest unmarried source of application downtime on Tinder. To evolve the condition of all of our caching system, we expected a sturdy and you may scalable provider.


I felt like rather very early one cache cluster management was a job that we wanted to conceptual regarding the builders as much that you could. I first believed playing with Amazon DynamoDB Accelerator (DAX) for our characteristics, however, sooner or later decided to play with ElastiCache to have Redis for a couple out-of explanations.

First and foremost, the app code currently uses Redis-situated caching and you will all of our current cache availability habits failed to lend DAX are a decrease-within the replacement particularly ElastiCache having Redis. Instance, some of the Redis nodes shop canned investigation away from several resource-of-truth data places, and now we found that we could perhaps not with ease arrange DAX having it objective.