Infrastructure Provisioning
Detailed overview of the various performance metrics captured by running RudderStack on various AWS configurations, to help you implement effective infrastructure provisioning
Last updated
Was this helpful?
Detailed overview of the various performance metrics captured by running RudderStack on various AWS configurations, to help you implement effective infrastructure provisioning
Last updated
Was this helpful?
This is a brief overview of the stats that we captured by running Backend on different AWS machine configurations that we hope gives a rough idea for users in making the decisions to provision infrastructure to host Rudder. All numbers capture below are by using metrics described in .
All tests are done using db.m4.xlarge Amazon RDS instance for hosting the postgres database
Machine
Load
Response Time (ms)
gateway.response_time
Throughput
gateway.write_key_requests
Dangling Tables
m4.2xLarge 8Core 32GB
2.5K/s
--
2.5K/s
No
m4.2xLarge 8Core 32GB
5K/s
--
3K/s
Yes
m4.2xLarge 8Core 32GB
3K/s
--
2.7K/s
No
m5.xLarge
4Core 16GB
2.5K/s
3
1.9K/s
No
m5.large
2Core 8GB
2.5K/s
4.2
1.7K/s
No
Backend migrates and drops tables that have a threshold of jobs processed. Gateway tables are backed up to object storage (S3, MinIO etc.) if configured by user. Dangling tables indicate tables are ready for drop at a rate greater than the rate at which tables are backed up to object storage. Concurrent uploads to object storage is in the roadmap for upcoming versions of Backend.
Machine
Gateway Throughput
gateway.write_key_requests
Throughput
processor.transformer_received
m4.2xLarge 8Core 32GB
2.7k/s
2.7K/s
m5.xLarge
4Core 16GB
1.9K/s
1.9K/s
m5.large
2Core 8GB
1.7K/s
1.6K/s
Transformer is a NodeJS/Koa server launced as cluster of node processes, processses count equal to the number of cores of the machine. Choosing an instance with lower number of core the number of cores in instance processor might reduce the throughput of transformer.
Machine
Gateway Throughput
gateway.write_key_requests
Throughput
batch_router.dest_successful_events
m4.2xLarge 8Core 32GB
2.7K/s
2.8K/s
m5.xLarge
4Core 16GB
1.9K/s
1.8K/s
m5.large
2Core 8GB
1.7K/s
1.6K/s
Below is an image captured in CloudWatch Metrics showing the captured stats
Rudder recommends using a database with at least 1TB allocated storage as there could be downtime to increase storage realtime depending on your database service provider.
If you want to dig deeper and figure out the right storage size, go through the following example. Following variables should be considered to come with a right storage size for your use case.
numSources
Total number of sources
2
numEventsPerSec
Number of events per sec for a given source
2500
avgGwEventSize
Event size that is captured at the gateway by Rudder
2.1 KB
gwEventOverhead
Size of extra metadata that Rudder stores at Gateway to process the event
300 B
numDests
Number of enabled destinations for a given source
3
avgRtEventSize
The payload size that needs to be sent by the router to the destination after applying transformations
1.2 KB
rtEventOverhead
Size of extra metadata that Rudder stores to process the event
300 B
gatewayStorage = numEventsPerSec * (avgGwEventSize + gwEventOverhead) routerStorage = numEventsPerSec * numDests * (avgRtEventSize + rtEventOverhead) totalStoragePerHour = 3600 * \sum_{firstSource}^{lastSource} (gatewayStorage + routerStorage)
In the above production example, after substituting the values, totalStoragePerHour adds up to 120 GB
Sample your peak load in production to estimate the storage requirements and substitute your values to get an estimate of the storage needed per 1 hour of data.
Event data and tables are ephemeral. In a happy path, we would have only a few minutes of event data being stored.
We recommend at least 10 hours worth of event storage computed above to gracefully handle destinations going down for a few hours.
If you want to prepare for a destination going for down for days, accommodate them into your storage capacity.
Rudder batches requests efficiently to write data. Under heavy load, backend can be configured (batchTimeoutInMS and maxBatchSize ) to batch more requests to limit concurrent connections to the database. If write latencies to the database are not in permissible thresholds, a new data set needs to be added i.e., backend server and database server.
Rudder reads the data back from the database at a constant rate. A sudden spike in user traffic will not result in more read DB requests.
Rudder does not cache aggressively and hence does not need huge amount of memory. Load tests were performed on 4 GB and 8 GB memory instances.
Rudder caches active user events by default to form configurable user sessions server side. The length of any user session can be configured with sessionThresholdInS and sessionThresholdEvents. Once a user's session is formed, that user events are cleared from the cache. If you don't need sessions, this can be disabled by setting processSessions to false.
numActiveUsers
Number of active users during a session (2 min) in your application during peak hours
10000
avgGwEventSize
Event size that is captured at the gateway by Rudder
2.1 KB
userEventsInThreshold
Number of user events in the given threshold i.e., 40 user events in 2 min
40
memoryNeeded = numActiveUsers * userEventsInThreshold * avgGwEventSize
Memory required in the above example would be 840 MB.
The memory estimate does not include the default RAM required for running the OS and the required processes.
Gateway Requests and Batch Router Throughputs
