[continuations]

A few questions:

1) >631851 tpmC

How many servers are needed to achieve this throughput?

2) >4 terabytes of unreplicated, frequently accessed data

4TB unreplicated data? Does that mean if a single node goes down you'll lose data (EDIT: I meant losing availability, not data)? That kinda ruins the whole point of having a distributed database.

3) If I'm reading the KV benchmaks correctly, it takes 5 nodes to achieve 100k tpm. That's 20k tpm per node. That's 333 tps per node. This is a 95% point read benchmark. Why is the tps (333 tps) so low? Is that normal?

4) How does CockroachDB compare to other distributed databases such as TiDB, FoundationDB, ScyllaDB?

[manigandham]

ScyllaDB is still the fastest at a key/value workload with per-query consistency settings and quorum reads/writes across multiple regions. If you need high-performance and low-latency, ScyllaDB wins. They are close to v3.0 which will have global secondary indexes and materialized views to improve data model flexibility. FoundationDB is also key/value but much lower-level and well proven for reliability. Don't have much experience with it and the latest release just introduced multi-regional capabilities, but the general tooling and documentation is still rough and it would take more effort to build a higher-level querying layer or client library.

TiDB is interesting, but missing more features from MySQL than CRDB is missing from PostgreSQL, so it's effective if you want sharding on mysql but will need a few more releases before it gets polished. Vitess and Citus are good options if you just want sharding on top of existing mysql or postgres with full query support within a shard. There's also Yugabyte which is multi-modal Redis/Cassandra/SQL offering with multi-regional capabilities.

CRDB is a great product with some of the easiest operations (although key management is a nightmare that they do not have a good plan for). It's fast enough for point-lookups and makes it easy to distribute and replicate your data across zones and regions. All nodes are part of a single cluster so read and write latencies will be high for global deployment, with the enterprise version having a workaround for local regional reads using pinned covering indexes. That works, but further lowers write performance.

It also has trouble with large transactions and the middle ground between OLTP and OLAP with heavy joins. Good choice if you need easy scalability and SQL interface over performance and complex queries.

[morgo]

Hi! I work for PingCAP, the company behind TiDB and come from previously working on MySQL.

The gap of features missing is documented here: https://www.pingcap.com/docs/sql/mysql-compatibility/

I would rate compatibility as actually pretty good: all but one SQL mode is supported (which is a feat in itself), and most of the SQL functions are supported.

There are some exceptions though, some which are addressable (missing functions) and some that are not (often a property of being an optimistic system).

We try to be as transparent as possible on this, which might be part of the reason why you feel there is a lot missing?

If you have specific examples, I would be happy to clarify. We also have a course designed for MySQL DBAs, which is designed to make the adoption easier: https://www.pingcap.com/tidb-academy

[manigandham]

Good to see the progress, I was looking at the roadmap page: https://github.com/pingcap/docs/blob/master/ROADMAP.md

Views and CTEs are probably the biggest missing pieces now.

[morgo]

The technical design for views was recently completed, and I expect to see them added soon :-)

Window functions & CTEs are only very recent features in MySQL 8.0 (TiDB is 5.7 compatible). None the less, they are important for HTAP workloads, and I'm looking forward to seeing them too.

[awoods187]

I'm the author of the post.

1. Between 90 and 135 16 vCPU nodes depending on cloud hardware 2. The cluster replicates this data three ways across all three nodes (so the cluster actually contains 12+tb of data) ensuring high availability. We intentionally reported the unreplicated number for clarity and comparison to TPC-C spec 3. Our graph is mislabeled. It should read transactions per second `tps`. Nice catch! 4. We can't comment on other database performance as they haven't release any TPC-C numbers.

[shenli3514]

“Between 90 and 135 16 vCPU nodes depending on cloud hardware ” How many nodes did you use in the CRDB 2.0 TPC-C 10k benchmark? Could I say that the "5x increment" is on the same hardware condition? Thanks!