Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Redis Cluster for Write Intensive Workloads

Redis Cluster for Write Intensive Workloads

NDC London 2021 (Remote)

When you are working with Redis for your write-intensive workloads, Redis Cluster is your friend. It gives you a built-in way to partition your data across instances to allow you to scale your writes without being bound to how much load a single instance can handle. However, data partitioning is always a challenge and Redis Cluster’s approach is no exception to that. At Deliveroo, we are using Redis Cluster in anger, for handing write intensive workloads (e.g. one use case has 10K writes per seconds, while simultaneously handling 300K reads per seconds). During the time we have been using Redis Cluster, we have gained learnings on how the basics of Redis sharding works, how upfront design choices can have tremendous impact on your performance to how resharding is handled both on the Redis Cluster side as well as through the Redis clients. In this session, I would like to share those invaluable learnings based on our battle-tested, real world experiences. At the end of the session, you should have a far better idea on how you can scale Redis for your write intensive workloads, and what type of surprises might be waiting for you.



Tugberk Ugurlu

January 29, 2021


  1. Redis Cluster for Write Intensive Workloads

  2. Hello, I’m Tugberk!

  3. None
  4. Sign up and get £10! https://roo.it/tugberku-dgfd

  5. careers.deliveroo.co.uk We’re Growing! Unique challenges, amazing people and great food!

  6. None
  7. Deliveroo Home Feed • Dense areas with a lot of

    restaurants • Making it hard for users to choose from the large selection • Each user's needs are different
  8. Jonny Brooks-Bartlett- How Deliveroo improved the ranking of restaurants -

    PyData London 2019 youtube.com/watch?v=bG95RmVOn0E • Already algorithmically ranking the Restaurant List through a rudimentary Linear Regression model • Desire to personalize this ranking for each user's needs • Predicting which restaurant a user is more likely to order from
  9. • Access to the aggregated user specific data from the

    ranking service on production • Costly to aggregate on production • Needs to be in-sync with the training pipeline and model serving. • Need a way to retrieve this data in optimum time for millions of users, while sustaining >1K rps, and keep this data up to date within a reasonable data consistency lag.
  10. None
  11. Canoe pipeline kicks in, aggregates the data for each user

    and serializes the data based on a protobuf format in 20 user bundles. Canoe Aggregating User Features Storing Protobuf{ed} Features From the Canoe pipeline, we pick up files which has protobuf data for 20 users and upload them to S3 S3 Queuing the Work For Each S3 File Mapping between S3 and SQS allows us to queue messages into SQS whenever there is a file upload on the S3 bucket SQS indexing each user features to Redis Cluster Lambda is kicked off by the event source mapping between SQS and the Lambda, which handles the Lambda Storing the data for O(1) access per user Redis Cluster is available to serve reads and writes with 3 primary shards and each having 1 replica Redis Cluster Reading the data from the Redis Cluster On production, we can access the user specific feature by issuing an O(1) query to redis cluster. Access
  12. • Data aggregation pipeline bundles 50 records per proto file,

    and uploads to a known S3 bucket • S3 object creation notification is enqueued to SQS • Lambda instances dequeues from from SQS, and writes to Redis Cluster
  13. None
  14. Allows you to scale the writes as well as the

    reads, which are good especially for unpredictable write workloads Allows you to increase the capacity with zero-downtime by adding new shard(s) and performing online resharding Reduces your blast radius, i.e. when a shard goes down, it only affects the portion of your data surface until a failover happens
  15. • Redis installation where data is sharded across multiple Redis

    nodes • These nodes still have the same capabilities as a normal Redis node, and they can have their own replica sets • Redis assigns "slot" ranges (a.k.a. hash slots) for each master node within the cluster
  16. tugberkugurlu/redis-cluster usage https://github.com/tugberkugurlu/redis-cluster

  17. • Redis comes with some out of the box commands

    to help you manage your cluster setup
  18. • For a given Redis key, the hash slot for

    that key is the result of CRC16(key) modulo 16384, where CRC16 here is the implementation of the CRC16 hash function • Redis clients can query which node is assigned to which slot range by using the CLUSTER SLOTS command
  19. None
  20. None
  21. None
  22. https://www.tugberkugurlu.com/archive/redis- cluster-benefits-of-sharding-and-how-it-works

  23. None
  24. • Gives a managed support for Redis Cluster mode (e.g.

    you don't need to worry about operational handling for resharding, failover, etc.) • Integrates well with our existing infrastructure stack at Deliveroo (e.g. AWS, Terraform, etc.)
  25. https://docs.aws.amazon.com/AmazonElastiCache/latest/red -ug/Replication.Redis-RedisCluster.html

  26. None
  27. None
  28. • READONLY command enables read queries for a connection to

    a Redis Cluster replica node. • RouteRandomly config option allows routing read-only commands to the random master or slave node. • These configurations allows us to distribute the read load across the master and all replicas in a random way at the cost of potentially increased data consistency gap.
  29. • Having tight timeouts allows us to reduce the impact

    of potential issues with the Redis to the rest of the application • If we know the expectations from the redis cluster in terms of response time, we can tune the timeout to fail early, allowing the rest of the application to keep executing in case of potential issues. • Timeout tuning is a half scientific and half finger in the air process...
  30. None
  31. None
  32. • Simple Redis Set command • The client knows which

    node to send this write request to thanks to its Redis Cluster knowledge
  33. • Simple Redis Get command • The contract between write

    and read side the is the userID • Checking on Redis error whether it's of type "redis.Nil" which indicates absence of the key.
  34. None
  35. None
  36. None
  37. None
  38. None
  39. None
  40. staurant features

  41. • Multi-command operations such as MGET can only succeed if

    all of the keys belong to same slot https://www.tugberkugurlu.com/archive/redis-cluster-benefits-of-sharding-and-how- it-works#hash-tags
  42. • Hash tags allow us to force certain keys to

    be stored in the same hash slot. • when the Redis key contains "{...}" pattern only the substring between { and } is hashed in order to obtain the hash slot. https://www.tugberkugurlu.com/archive/redis-cluster-benefits-of-sharding-and-how- it-works#hash-tags
  43. • None of the access pattern needs was requiring us

    to go across city boundary • Therefore, used City ID as the hash tag value
  44. • Same as the write side, we use City ID

    as the hash tag here to influence the shard selection to route us to the same node • Bundling all Redis Get commands within a single TCP connection to improve the performance by saving from the round trip • Pipeline requests run in order but they are not blocking other connections unlike MGET
  45. • around ~850-1K queries per second • ~9.72ms max p95

    latency for entire pipeline query
  46. None
  47. None
  48. • Increasing the number of node groups for your Elasticache

    Cluster will kick off an online resharding operation • This will inherit the same number of replications as the other node groups
  49. None
  50. • You can increase/decrease the replica count independent of the

    shard count • Note that there was a bug on Terraform regarding this but it has been fixed, see github.com/hashicorp/terraform-provider-aw s/issues/6184 https://docs.aws.amazon.com/AmazonElastiCache/latest /APIReference/API_IncreaseReplicaCount.html
  51. None
  52. None
  53. https://docs.aws.amazon.com/AmazonElastiCache/latest/red-ug/AutoFailover.html#auto-failover-test

  54. None
  55. 1 4 2 5 3 6

  56. 56 Software Engineer - Mid, Senior, Staff-level Engineering Manager Senior

    Software Engineer, Infrastructure Machine Learning Engineer - Mid, Senior, Staff-level Data Engineer Data Scientist - Mid, Senior, Staff-level Data Science Manager Locations: London, Remote UK, Remote Poland See the complete list at https://careers.deliveroo.co.uk/ !
  57. None
  58. None