asyncio and blocking - an update

Chris Angelico rosuav at gmail.com
Thu Feb 11 00:57:49 EST 2016


On Thu, Feb 11, 2016 at 4:45 PM, Frank Millman <frank at chagford.com> wrote:
> I have come up with a plan that seems to provide a solution.
>
> Instead of 'putting' one row at a time, let the database handler build up a
> block of rows, and then 'put' the block.
>
> I tried a block of 10, and it ran a lot faster. I increased it to 50, and it
> ran faster again. I tried 100 and there was not much improvement, so 50
> seems like an optimum number. The speed is now only slightly slower than
> run_in_executor(), and it is more truly asynchronous.

Something worth checking would be real-world database performance
metrics: what's time-to-first-row versus time-to-subsequent-rows? When
you submit a query, the server first has to parse it and check for
errors, then do all its optimization and stuff, and figure out an
access plan. Then it goes and fetches stuff. If your query is a simple
"select * from tablename" on a huge table, then it's entirely possible
that you save a lot of time by fetching subsequent rows
asynchronously; but if there's an ORDER BY that can't be resolved from
an index (maybe involving a table join or a non-optimizable function
call), the database might have to read everything from the disk before
it can return a single row, so the time from first row to last row is
a tiny fraction of the time to first row. What are your actual
real-world queries like?

Most likely, the database is reading rows in pages, and you'll have no
way of predicting how many usable result rows are on any page. I
wonder, is there any way you can actually operate this the other way
around? Presumably the database is connected to you via a socket of
some sort (TCP if you're going over a network, maybe a Unix socket for
local connections), which you could select() on same as any other; if
you can react to the arrival of more rows, that might be the easiest
solution. That would probably require a dedicated async database
library, though.

ChrisA


More information about the Python-list mailing list