issue_comments

I did two tests: one with 1000 5-30mb DBs and a second with 20 multi gig DBs. For the second, I created them like so: for i in {1..20}; do sqlite-generate db$i.db --tables ${i}00 --rows 100,2000 --columns 5,100 --pks 0 --fks 0; done

This was for deciding whether to use lots of small DBs or to group things into a smaller number of bigger DBs. The second strategy wins.

By simply persisting the _internal DB to disk, I was able to avoid most of the performance issues I was experiencing previously. (To do this, I changed the datasette/internal_db.py:init_internal_db creates to if not exists, and changed the _internal DB instantiation in datasette/app.py:Datasette.__init__ to a path with is_mutable=True.) Super rough, but the pages now load so I can continue testing ideas.

10-20 minutes to populate _internal! How many databases and tables is that for?

I may have to rethink the _internal mechanism entirely. One possible alternative would be for the Datasette homepage to just show a list of available databases (maybe only if there are more than X connected) and then load in their metadata only the first time they are accessed.

I need to get my own stress testing rig setup for this.

Upon first stab, there's two issues here: - DB/table/row counts (as discussed above). This isn't too bad if the DBs are actually above the MAX limit check. - Populating the internal DB. On first load of a giant set of DBs, it can take 10-20 mins to populate. By altering datasette and persisting the internal DB to disk, this problem is vastly improved, but I'm sure this will cause problems elsewhere.

I'm running into issues with this as well. All other pages seem to work with lots of DBs except the home page, which absolutely tanks. Would be willing to put some work into this, if there's been any kind of progress on concepts on how this ought to work.

I'm seeing the problem on database page. Index page and table page runs quite fast.

• Tables have <10 columns (id, url, title, body_html, date, author, meta (for keeping unstructured json)). I've added index on date columns (using sqlite-utils) in addition to the index present on id columns.
• All tables have FTS enabled on text and varchar columns (title, body_html etc) to speed up searching.
• There are couple of tables related with foreign keys (think a thread in a forum and posts in that thread, related with thread_id)

I've just realized Datasette tries to count hidden tables too. There are 5 visible tables, 25 hidden tables, which I haven't realize earlier to consider their effect. I've turned off counting for hidden tables to see if it has any effect.

What's the point of counting FTS tables?

I tried setting cache_size_kb=0 then cache_size_kb=100000, still getting this behavior. I even changed Database::table_counts and lowered time limit to 1

py table_count = ( await self.execute( "select count(*) from [{}]".format(table), custom_time_limit=1, ) ).rows[0][0] counts[table] = table_count

I feel like 10 seconds is a magic number, like a processing timeout and datasette gives up and returns the page. Index page loads instantly, table page, query page, as well. But when I return to database page after some time, it loads in 10s.

EDIT:

It's always like 10 + 0.3s, like 10s wait and timeout then 300ms to render the page

I wonder if this is a SQLite caching issue then?

Datasette has a configuration option for this but I haven't spent much time experimenting with it so I don't know how much of an impact it can have: https://datasette.readthedocs.io/en/stable/config.html#cache-size-kb

The issue seems to appear sporadically, like when I return to database page after a while, during which some records have been added to the database.

I've just visited database, page first visit took ~10s, consecutive visits took 0.3s.

I'm chunking inserts at 100 at a time right now: https://github.com/simonw/sqlite-utils/blob/4d9a3204361d956440307a57bd18c829a15861db/sqlite_utils/db.py#L1030

I think the performance is more down to using Faker to create the test data - generating millions of entirely fake, randomized records takes a fair bit of time.

Hmm, not seeing the problem now.
I've removed the commented out sections in database.py and restarted the process. Database page now loads in <250ms.

I have couple of workers that check some pages regularly and scrape new content and save to the DB. Could it be that datasette tries to recount tables every time database size changes? Normally it keeps a count cache, but as DB gets updated so often (new content every 5 min or so) it's practically recounting every time I go to the database page?

EDIT: It turns out it doesn't hold cache with mutable databases.

I'll update the issue with more findings and a better way to reproduce the problem if I encounter it again.

sqlite-generate many-cols.db --tables 2 --rows 200000 --columns 50

Looks like that will take 35 minutes to run (it's not a particularly fast tool).

Try chunking write operations into batches every 1000 records or so.

You said "200k+, 50+ rows in a couple of tables" - does that mean 50+ columns? I'll try with larger numbers of columns and see what difference that makes.

Ah that was a typo, I meant 50k.

I think I should mention that having FTS on all tables mean I have 5 visible, 25 hidden (FTS) tables displayed on database page.

Just to check: are you seeing the problem on this page: https://latest.datasette.io/fixtures (the database page) - or this page (the table page): https://latest.datasette.io/fixtures/compound_three_primary_keys

If it's the table page then the problem may well be #862.

sqlite-generate many-cols.db --tables 2 --rows 200000 --columns 50

Looks like that will take 35 minutes to run (it's not a particularly fast tool).

I generated a 600MB database using sqlite-generate just now - with 100 tables at 100,00 rows and 3 tables at 1,000,000 rows - and performance of the database page was fine, 250ms.

Those tables only had 4 columns each though.

You said "200k+, 50+ rows in a couple of tables" - does that mean 50+ columns? I'll try with larger numbers of columns and see what difference that makes.

I'm not sure if table counts are to blame. There shouldn't be a ~3 orders of magnitude difference.

```fish user@klein /a/w/scrapyard (master)> set sql "select count() from table_1; select count() from table_2; select count(*) from table_3;" user@klein /a/w/scrapyard (master)> time sqlite3 scrapyard.db "$sql" 187489 46492 2229

Executed in 25.57 millis fish external usr time 3.55 millis 0.00 micros 3.55 millis sys time 22.42 millis 1123.00 micros 21.30 millis ```

but not letting datasette count the tables definitely helps.

I'll write a little script which generates a 300MB SQLite file with a bunch of tables with lots of randomly generated rows in to help test this.

Having a tool like that which can generate larger databases with different gnarly performance characteristics will be useful for other performance work too.

This makes a lot of sense. I implemented the mechanism for the index page because I have my own instance of Datasette that was running slow, but it had a dozen database files attached to it. I've not run into this with a single giant database file but it absolutely makes sense that the same optimization would be necessary for the database page there too.

Oops, the same method is called from both index and database pages. But removing select count queries speed up the page load quite a bit.