I've extensively used Dynamo (internally at Amazon and externally) and even founded a DB startup with it at it's core. Boiling down scalability of Postgres vs Dynamo as it's written in blog is a bit terse. Dynamo scales writes horizontally with the keyspace, forever. Postgres simply can't, and no number of layers between the machines and the developer changes that. Sharding, pooling, Citus are all layered on top of an engine where a given row's writes still land on one
primary.
If you know your access patterns really well and they are non-relational, then you can design the best possible tables for dyanmoDB. In such a case DynamoDB works and scales amazingly. Ofc, you cannot do multi table relationships etc shoehorning a relational scheme onto DynamoDB does not work.
Dynamo is a fundamentally different DB to Postgres. If your problem fits into the dynamo approach (I'd argue that more problems do), then you should be using it. No all problems fit, though.
Agreed, my critique was about how the article frames scalability. I've yet to see an OLTP problem that can't live in something like Dynamo. KV can model anything if you put in the work, the question is how much modeling discipline you trade for the scale, and in my experience the up front work is always worth it. Most of the time operational issues are swept under the rug and not consider tech debt.
Take for example AuroraDB: the sheer engineering it took to make SQL do scalable OLTP at all tells you how much that flexibility actually costs to keep.
Upfront modeling work is always worth it, but that only holds if you actually know your access patterns upfront. Most teams don’t, especially early on.
Curious how the DB startup with Dynamo at its core went. We use it heavily. The primary tricky thing for us at the moment is aligning pricing with workload value.
We obsessed over optimizations and pushing the apis to the limits of how we could pack it.
So much so, we re-wrote the DynamoSDK to squeeze out more optimizations so we could be the same cost even though we were a layer in front of dynamo. We used key encoding and other various technique as well as managed capacity (on demand vs reserved) to transparently optimize workloads for price. In our experience we saw dramatic gains vs just vanilla SDK usage.
If you're curious, here was the marketing website, but we're now part of Databricks: https://stately.cloud/
Interesting! We interact with the low-level APIs too vs the SDK, also: an IO scheduler for request batching and conn management, request hedging, full MVCC transactions, etc. We store raw bytes in DDB and manage schema/etc elsewhere. Curious if there is other low-hanging fruit, or not so low, you found that we haven't discovered yet.
Not by itself if it's naive, but if it's able to assess target health and avoid degraded instances then it becomes a component in HA, the other being integrating an orchestrator for gracious recovery.
> PgDog does not detect primary failure and will not call pg_promote(). It is expected that the databases are managed externally by another tool, like Patroni or AWS RDS, which handle replica promotion.
HA has to be all the way through, in which case you might not need a load balancer because each client already connects to a separate server. If you do, then you can have one load balancer per client machine.
I tend towards using key-value databases as I find them general purpose enough while being much more robust. I'm not married to any one in particular, depends on the requirements.