Pg8000 – Pure-Python PostgreSQL driver

Scaling for one. If 95% of your workload is non-transactional reads, why introduce transactions?

If your writes are single table and don't require transactional safety, why introduce transactions?

If you need complex, multiple-insert or update writes, transactions are great, but that can also be accomplished in a stored proc.

In a high-velocity environment, avoiding transactions are a great way to get more throughput out of your database server with lower latencies.

Combine Postgres with pgBouncer in statement mode and you can scale it pretty dramatically.

I would instead look at it as turning autocommit "off" is more like turning "automatic transaction begin" "on", which is just a strange semantic. I could appreciate a mechanism that insisted "transactions must always be explicit", but that isn't what autocommit controls, on or off. (You also wouldn't want such unless you were assured that you could put the transaction boundaries inside of multi-statement statements, as you really really want to minimize round trips there as otherwise you can end up--even on the same machine--drastically harming your lock contention by having transactions that last thousands of times longer than the single statement they wrapped... at which point the API is going to get awkward trying to enforce the transaction boundaries... autocommit is essentially just what you should want as the solution to all of this.)

I think this illustrates my point ... but I'm not 100% sure what you mean by "autocommit mode" in your comment! You said "autocommit feature in DB drivers" but I think you mean in the DB itself by the sounds of it.

AFAIK autocommit mode just applies if you don't explicitly open a transaction, so what would you propose instead? That a transaction is silently opened but not committed (potentially leaving a dangling transaction open for a long time after SELECT statement) (like Python's odd default behaviour)? Or an immediate error?

I can certainly see that if you are indeed wrapping every individual statement in their own BEGIN/COMMIT pairs that you'd wonder why the database can't help you for such simple use cases. You can still selectively open an explicit transaction on those statements that do need to be grouped together, so I don't see that you've given up any control.

There are a lot of people who use databases without really understanding anything about them. They mostly won't understand why it is important to control time and execute multiple statements at the same logical moment in database world. That is a difficult concept to intuit naturally, people need to be told and they often don't have time or a teller.

Databases that only do the technically-correct thing are weeded out because they lose the big supportive cloud of novices that experts grow out of. So the good databases support the intuitive way and the the correct way of manipulating data.

There are more computing concepts to learn than there are hours to learn it in. People use systems they don't understand and systems need to support that.

Up until 10ish years ago, the most popular SQL DB didn't support transactions in the default setup. So lots of apps got written without relying on having transactions, either for compatibility reasons or because the devs grew up on said nontransactional DB. So autocommit was an easy way to emulate transactionless SQL on the rest of the DBs.

In practice autocommit works very well. In many workloads most transactions consist of just one query. Additional if you have autocommit it is harder to accidentally create a long running transaction which is a good thing since long running transactions can really tank your performance.

To all those wondering _why_ that is the recommendation, I did some digging and the answer is "the binary package is compiled against one version of libssl, and if your system happens to use a different one, in some weird edge cases it will lead to a segfault"

I don't know, that's what the developers say (last paragraph): https://pypi.org/project/psycopg2/

More details of my use case:

The programs were ETLs that are generating big REPLACE INTO statements.

Different databases and drivers do prepared statements in different ways. Postgres, for example, can receive the statement and the parameters and do the interpolation in the server. The params still need to be serial used by the client and the template isn’t reused, so it’s no big perf win.

MySQL, on the other hand, needs the driver to do the interpolation. The server only knows how to execute sql strings.

It turns out it is far better to prepare the sql in pure python where pypy knows all about types etc than to have the c MySQL driver do it.

Exactly the same thing plays out in java. The MySQL java drivers are staggeringly inefficient at large prepared statements, and in my profiling massive amounts of time is spent in java functions in the MySQL driver that are trying to work out where the next question-mark is. Sadly the drivers are the poo they are. Iirc the mariadb java drivers are rather more optimized.

> That’s surprising if your app spent most of the time in database calls, since 1) native code is still faster than pypy and 2) most of the time spent in database libraries is spent waiting for the database to return

I would argue a speedup (though maybe not a “massive” one) from pure python over a native extension is not that surprising, specifically because of (2), which makes it possible for the runtime-extension interface overhead to be more significant than (1). Native extensions tend to give a speed boost for computationally intensive tasks, but can’t really wait any faster, and aren’t free of overhead.