Would love to hear from someone who've used them in production.
After trying some different options it became clear that Cockroach was the best option by far. The strongly consistent nature of Cockroach made it a good choice for a SoR like whistleblowing unlike say Mongo. The ease of deployment and the stability of the Raft consensus protocol for managing ranges of data was better than expected.
For me having done a global rollout of it, I’d recommend it with very few caveats. You do need to train developers on what queries it can and cannot support, since it does have Postgres client support but obviously cannot handle the full range of queries. You also need to have lots of alerting around the Raft log, since that is your bread and butter in terms of “is this working or not”. However in aggressive testing, removing nodes with no warning under load, simulating pretty severe networking problems, etc, it has held up remarkably well.
Check out Aerospike. It's the most robust and truly "zero effort scale" real time data platform I've ever seen. Properly configured and used, it whups everything in terms of TPS and performance/machine aka TCO. By a huge, huge, factor.
Downside is ... everything else. :/ Less polished developer experience, docs, &c. There are so many huge customers, but very few who are talking about it. Requires more of a time investment -- because you're learning a new platform, not using a drop in replacement for what you already know. But the reward is performance beyond anything else.
(the VLDB 2016 paper is a good high level overview of the architecture)
If you do check it out, be sure to use version 6.1 or later, in particular see what can be done combining secondary indexes with container data types.... A, uh, birdie told me that. :P
edit: Same birdie told me Aerospike has customers who replaced CockroachDB, when they found out it has trouble scaling past a point.
I really hope they can get their act together in terms of communication and marketing -- it's sadly "the most advanced solution you've never heard of".
Launched on a t3.micro (1GB RAM) - same.
Launched on a t3.small (2GB RAM) - finally it works and doesn't crash.
So, min 2GB RAM to serve about 10 (as in: ten, not thousands) records queried every other second.
Being a die hard fan of Postgresql for more than 15 years, I think time has come for me to select CockroachDB as new default.
I miss Table Partitioning (only available in enterprise edition ) and Full text search (yet to be built). Also minimum hardware I had to employ was quad core and 8 gigs of RAM.
But other than that everything is so smooth including installation, online documentation. It made me lazy.
I have used an alternative.