Hacker News Reader

Hacker News

Comment by rawnlq | Hacker News Reader

Full thread

rawnlq

Uber switched from mysql to postgres in 2013:

https://www.yumpu.com/en/document/view/53683323/migrating-ub...

but switched back to mysql in 2016 which generated a lot of discussion:

https://news.ycombinator.com/item?id=12166585

There were a lot of followup posts afterwards arguing both sides if you search for them:

https://news.ycombinator.com/item?id=12216680

View on HN

There's an excellent analysis of the Uber's post here: http://use-the-index-luke.com/blog/2016-07-29/on-ubers-choic...

It's strange to me everyone touts Postgres but when you finally scale to a medium to large architecture HN has a handful of articles on great start-ups (now full fledge companies) switching FROM Postgres to MySQL. It's more efficient, atm, especially when accounting for labor. I wish the HN community would see that.

There are more MySQL experts out there than there are PostgreSQL experts (though the number is still probably under 10,000 worldwide). MySQL replication and scaling is a well understood solution, and thoroughly battle tested. There are dozens of good solutions out there for handling near-instantaneous failovers, using little more than the tools provided by MySQL.

It has a lot of problems, but when you need to scale, you can do so much more affordably with MySQL.

In Uber's case they wanted a dumb K/V store, not a full RDBMS. MySQL is a better choice here, InnoDB is a great K/V store, but PostgreSQL is a better relational database assuming you're using the features it provides.

Regarding InnoDB being a great K/V store, you made a similar claim yesterday, but you haven't really explained your rationale in either thread. Care to do so?

EDIT: I'm referring to https://news.ycombinator.com/item?id=14067406 ; my reply to that comment included more specific questions. Unsure of why receiving downvotes here; being somewhat of an InnoDB expert myself, I am legitimately curious why the parent keeps making assertions about InnoDB being a K/V store.

That's because InnoDB IS a key-value store, the physical on disk structure is such that rows are stored on disk by their primary key, and that's how MySQL is designed. Postresql's heap files are organized much differently, as such it performs worse in some cases like Uber's.

That's a clustered index, but it does not inherently make it a key/value store. It is common practice in SQL Server to use a clustered index for tables; would you also think of that as a key/value store? Most database people don't describe clustered indexes in this way.

Also, you're describing how InnoDB is designed, not how MySQL is designed. Furthermore, you can effectively get a heap in InnoDB if a table has no PK and no unique indexes. InnoDB will internally use a hidden auto-incremental value in this case for the primary lookup, which more or less effectively acts as a row pointer.

I thought the reason for migrating over was for MemSQL?

Here is a video "Real Time Analytics at UBER Scale": https://www.youtube.com/watch?v=XSXvPNEykb4

Reply on news.ycombinator.com