Hi Rob,
Is there significant read access (iostat: r/s) during these slow writes ? If
yes, it might be due to a small cache, requiring the database to read (index)
pages before updating them.
And is the data you're adding in both databases (large/small) added at the end
of the table using the autoincrement, or do you insert some items in the middle
? I'm not a SQLite performance expert, but in other databases performance can
be very different because in the former case fewer pages must be updated.
Microsoft SQL Server has something called "fill factor", basically the max
percentage of an index page that is used during initial fill, which helps
avoiding too many page shuffling in the index when extra items are added.
Disadvantage: it makes DBAs argue endlessly about the best fill factor ;-)
Maybe there's something similar possible in SQLite but I couldn't find a pragma
for this.
Oh, and do both databases have the same page size, preferably 4K ?
Regards,
Droedel
On Mon, Jul 30, 2018, at 12:51, Rob Willett wrote:
> Droedel,
>
> Thanks for the comprehensive reply. We have actually done all of this.
>
> The system has been running for 2-3 years and we have taken the
> opportunity to try and prune the database from 60GB down to 600MB.
> Currently the live system is working OK with a 60GB database, but our
> test system (which is smaller) is struggling with 600MB.
>
> The system has a restriction of IOPS as it's a Virtual Private Server.
> Technically it's running Ubuntu 16.04 under OpenVZ. We can get 69MB/sec
> with a disk to disk copy, which isn't brilliant if we had sustained disk
> traffic which we don't.
>
> We log approx 600 - 800 items of around 3-5K every five minutes. These
> 600-800 items are mainly an insert into a single table, there are other
> things happening as well, but this is the bulk of the work. We can see
> that the -wal files grow a small amount (4-5MB) just before the commit.
> It then takes 7 seconds to execute the commit. This is the bit that
> we're struggling with. We know we can get circa 70MB/sec data
> throughput, so this should take a fraction of a second. Now SQLite needs
> to work out which pages to commit so thats a little longer, but we know
> SQLite is fast, so that shouldn't take 7 seconds on the small database
> as it doesn't take that long on the large 60GB database. Thats the
> puzzling bit, the large database is quick, the small one slow.
>
> We have no logging turned on, we can turn SQL logging on at the DBI
> level but that turns a 20 sec run into a 2-3 minute run as it captures
> everything :) Nothing in the log files gives us any concern (apart from
> the really long commit time). Simon Slavin suggested dropping the
> indexes which we did, that turned the commit into a fast commit, so its
> something to do with the indexes but we can't see what.
>
> What we are now doing is going back to the very beginning:
>
> 1. We built a replacement system yesterday with 8GB memory and 8 cores
> and 150GB disk space. Its virtualised (ESXI) but under our control.
> 2. We've installed a copy of the old 60GB database on the new system.
> 3. We're going to benchmark the new system over a couple of thousand
> runs to see what the average time is.
> 4. We'll then work our way through the deduping of the database step by
> step to see when the commit time blow up. This will take a few days as
> working out the duplications of 200,000,000 rows isn't that quick :) As
> we found out, dropping a very large table is really, really, really
> slow.
> 5. We'll apply some of the ideas that people have suggested since
> yesterday to see if they work, but I'm keen that we have a repeatable
> problem that we solve rather than we use a scatter gun approach to
> fixing it. We think SQLite is well written so we figure the problem is
> ours to solve rather than simply blaming the software.
>
>
> Thanks
>
> Rob
>
> On 30 Jul 2018, at 11:11, Droedel wrote:
>
> > Hi,
> >
> > When having bad performance, I usually first try to find out if the
> > slowness is due to disk througput (sequential), slow random access or
> > something else. In Linux, try "iostat -xtc 5". Do this with and
> > without your application writing to disk.
> >
> > If you see high CPU %iowait and high %util on your disk, then disk is
> > the bottleneck. If not: start profiling / analyzing other bottlenecks
> > (CPU / network / ...)
> >
> > If the disk throughput (wMB/s) is close to your normal sequential
> > throughput (69 MB/s): try to write less data or get a faster disk.
> > If the disk troughput is low, but high numbers of writes (w/s):
> > there's too much seeking / too many small writes to your disk. Page
> > cache too small ? Checkpointing too often ?
> >
> > Sometimes this kind of problems is caused by other applications
> > (logging / ...) causing too much baseload. %util should be low when
> > your application isn't running.
> >
> > Just my 2 cents.
> >
> > Kind regards,
> >
> > Droedel
> >
> >
> > On Sun, Jul 29, 2018, at 10:14, Rob Willett wrote:
> >> Hi,
> >>
> >> Background
> >>
> >> We've been refactoring our database to reduce the size of it. Through
> >> some simple logic we've managed to pull out 99% of the data to reduce
> >> the size from 51GB down to approx 600MB. This logic has been to
> >> remove
> >> rows that are almost the same but not quite identical. As with all
> >> things, the thinking was the difficult bit, the execution somewhat
> >> easier.
> >>
> >> As part of the testing we've been doing, we've now hit on an odd and
> >> weird problem to do with the COMMIT statement. A commit of a few
> >> hundred
> >> (circa 600-800) rows takes approx 7 seconds whereas before we never
> >> even
> >> noticed it, though we now know it was two seconds before. Each row is
> >> probably 1-2K of data, so its not very much at all.
> >>
> >> Details of what we have tried:
> >>
> >> 1. We've turned synchronous on and off
> >>
> >> PRAGMA synchronous=ON
> >>
> >> and thats not made any difference.
> >>
> >> 2. We are using and have been using WAL mode for years.
> >>
> >> PRAGMA journal_mode;
> >> journal_mode
> >> wal
> >>
> >> 3. We've tested that the server copies OK, we get a consistent
> >> 69MB/sec.
> >> This is not as fast we would like, but it's the same across all our
> >> virtual servers.
> >>
> >> 4. We've tested the commit on our existing 60GB database and it takes
> >> 2
> >> seconds, which is longer than we thought it would be. The server for
> >> the
> >> 60GB database is a large VPS with 8GB/8 cores and runs Ubuntu 14.04.
> >> The
> >> server we are testing on is a 2GB/2 core test server running Ubuntu
> >> 16.04. Whilst the test server is smaller, we wouldn't expect it to
> >> take
> >> 3 times longer to do a commit.
> >>
> >> 5. The code is identical across the servers. We are running Perl and
> >> the
> >> DBI module. The code for doing a commit in Perl::DBI is
> >> $dbh->do("COMMIT");
> >>
> >> We are getting the expected performance elsewhere on the system
> >> and
> >> in the code. It's just the commit that is taking a long time.
> >>
> >> 6. The code we are committing is adding 600-800 lines to a table that
> >> used to be 200,000,000 rows in size. It's now 400,000 lines in size.
> >> We
> >> are wondering if the deletion of the lines has had an impact we
> >> didn't
> >> expect. We have vacuumed and analysed the database.
> >>
> >> The schema for the table we insert into is
> >>
> >> CREATE TABLE IF NOT EXISTS "Disruptions" (
> >> "id" INTEGER NOT NULL PRIMARY KEY AUTOINCREMENT,
> >> "version" integer NOT NULL,
> >> "Disruption_id" INTEGER NOT NULL,
> >> "status" integer NOT NULL,
> >> "severity" integer NOT NULL,
> >> "levelOfInterest" integer NOT NULL,
> >> "category" integer NOT NULL,
> >> "subCategory" integer NOT NULL,
> >> "startTime" TEXT NOT NULL,
> >> "endTime" text NOT NULL,
> >> "location" integer NOT NULL,
> >> "corridor" integer NOT NULL,
> >> "comments" integer NOT NULL,
> >> "currentUpdate" integer NOT NULL,
> >> "remarkTime" TEXT NOT NULL,
> >> "lastModTime" TEXT NOT NULL,
> >> "CauseAreaPointX" real NOT NULL,
> >> "CauseAreaPointY" real NOT NULL,
> >> "Direction" TEXT
> >> );
> >> CREATE INDEX "Disruptions_Idx1" ON Disruptions ("location" COLLATE
> >> NOCASE ASC, "corridor" COLLATE NOCASE ASC, "status" COLLATE NOCASE
> >> ASC,
> >> "category" COLLATE NOCASE ASC, "severity" COLLATE NOCASE ASC,
> >> "levelOfInterest" COLLATE NOCASE ASC, "version" COLLATE NOCASE ASC,
> >> "subCategory" COLLATE NOCASE ASC, "startTime" COLLATE NOCASE ASC);
> >> CREATE UNIQUE INDEX "Disruptions_Idx2" ON Disruptions
> >> ("Disruption_id"
> >> COLLATE NOCASE ASC, "version" COLLATE NOCASE ASC, "category" COLLATE
> >> NOCASE ASC, "subCategory" COLLATE NOCASE ASC);
> >> CREATE UNIQUE INDEX "Disruptions_Idx3" ON Disruptions ("version"
> >> COLLATE
> >> NOCASE ASC, "Disruption_id" COLLATE NOCASE ASC, "location" COLLATE
> >> NOCASE ASC);
> >> CREATE INDEX Disruptions_Idx5 ON Disruptions ("status",
> >> "Disruption_id",
> >> "Severity", "levelOfInterest", "category", "subCategory", "version");
> >>
> >> We have checked that this schema is consistent across the databases.
> >>
> >> We're about to recreate the table to see if that makes a difference.
> >>
> >> Any help or advice welcomed.
> >>
> >> Thanks
> >>
> >> Rob
> >> _______________________________________________
> >> sqlite-users mailing list
> >> sqlite-users@mailinglists.sqlite.org
> >> http://mailinglists.sqlite.org/cgi-bin/mailman/listinfo/sqlite-users
> > _______________________________________________
> > sqlite-users mailing list
> > sqlite-users@mailinglists.sqlite.org
> > http://mailinglists.sqlite.org/cgi-bin/mailman/listinfo/sqlite-users
> _______________________________________________
> sqlite-users mailing list
> sqlite-users@mailinglists.sqlite.org
> http://mailinglists.sqlite.org/cgi-bin/mailman/listinfo/sqlite-users
_______________________________________________
sqlite-users mailing list
sqlite-users@mailinglists.sqlite.org
http://mailinglists.sqlite.org/cgi-bin/mailman/listinfo/sqlite-users
