Benchmarks and comparisons for these two NoSQL databases exist a plenty. But in all the blogs and whitepapers I found the use-case was quite different from ours, we were trying to find a tool to queue Icinga 2 events and speed up config updates. The queue would use Redis Pub/Sub or MongoDBs capped collections + tailable cursors. From an implementation standpoint MongoDB looks better, it offers good libraries for most programming languages and the document-based approach allows for filtering based on individual attributes while values in Redis are just opaque blobs. Redis Pub/Sub model also has the problem that it loses its queued items when restarted.
From these naive benchmarks (PHP scripts run with time), it looks like MongoDB likes to take their time with creating objects.
Inserting 250.000 objects:

  • Redis: 11.1s
  • MongoDB: InsertOne: 53.3s
  • MongoDB: InsertMany: 25.4s

Inserting 250.000 objects into list capped to 50.000

  • Redis: 12s
  • MongoDB: InsertOne: 57.6s

Redis speed advantage comes from keeping created objects in memory, while MongoDB writes them to the hard drive right away. This is a huge advantage for Redis, as we rotate our events from the queue, writing them to disk would be a waste of io time. But sadly MongoDB does not offer the possibility of keeping the data volatile in memory.
Another concern of ours is security, Redis does not offer any security features and for many application this is not a problem, but in our case, this means we have to finds ways to make it secure. There have been many attempts to do so, e.g using tools like stunnel, but it’s still an additional topic that would need to be tackled. MongoDB on the other hand offers ACLs on document level and native SSL, which is great! Yet is has been rightfully criticized for its insecure default configurations, which resulted in a flood of hacked MongoDB instances earlier this year.
Conclusion: We will have to investigate further. Especially integration and security will need a closer look. MongoDBs performance seems lackluster, but may just be enough as our tests may have been much bigger than what we will be met with in a real setup.