Sqlite3: How to interrupt a long running update without roll back? - sqlite

Question

I have a long running multirow update such as:
UPDATE T set C1 = calculation(C2) where C1 is NULL
If table is large this update may take many seconds or even minutes.
During this time all other queries on this table fail with "database is locked"
after connection timeout expires (currently my timeout is 5 seconds).
I would like to stop this update query after, say, 3 seconds,
then restart it. Hopefully, after several restarts entire table will be updated.
Another option is to stop this update query before making any other request
(this will require inter-process cooperation, but it may be doable).
But I cannot find a way to stop update query without rolling back
all previously updated records.
I tried calling interrupt and returning non-0 from progress_handler.
Both these approaches abort the update command
and roll back all the changes.
So, it appears that sqlite treats this update as a transaction,
which does not make much sense in this case because all rows are independent.
But I cannot start a new transaction for each row, can I?
If interrupt and progress_handler cannot help me, what else I can do?
I also tried UPDATE with LIMIT and also WHERE custom_condition(C1).
These approaches do allow me to terminate update earlier,
but they are significantly slower than regular update
and they cannot terminate the query at specific time
(before another connection timeout expires).
Any other ideas?
This multirow update is such a common operation
that, I hope, other people have a good solution for it.

Answer 1

So, it appears that sqlite treats this update as a transaction, which does not make much sense in this case because all rows are independent.
No, that actually makes perfect sense, because you're not executing multiple, independent updates. You're executing a single update statement. The fine manual says
No changes can be made to the database except within a transaction.
Any command that changes the database (basically, any SQL command
other than SELECT) will automatically start a transaction if one is
not already in effect. Automatically started transactions are
committed when the last query finishes.
If you can determine the range of keys involved, you can execute multiple update statements. For example, if a key is an integer, and you determine the range to be from 1 to 1,000,000, you can write code to execute this series of updates.
begin transaction;
UPDATE T set C1 = calculation(C2)
where C1 is NULL and your_key between 1 and 100000;
commit;
begin transaction;
UPDATE T set C1 = calculation(C2)
where C1 is NULL and your_key between 100001 and 200000;
commit;
Other possibilities . . .
You can sleep for a bit between transactions to give other queries a chance to execute.
You can also time execution using application code, and calculate a best guess at range values that will avoid timeouts and still give good performance.
You can select the keys for the rows that will be updated, and use their values to optimize the range of keys.
In my experience, it's unusual to treat updates this way, but it sounds like it fits your application.
But I cannot start a new transaction for each row, can I?
Well, you can, but it's probably not going to help. It's essentially the same as the method above, using a single key instead of a range. I wouldn't fire you for testing that, though.
On my desktop, I can insert 100k rows in 1.455 seconds, and update 100k rows with a simple calculation in 420 ms. If you're running on a phone, that's probably not relevant.

Answer 2

You mentioned poor performance with LIMIT. Do you have a lastupdated column with an index on it? At the top of your procedure you would get the COMMENCED_DATETIME and use it for every batch in the run:
update foo
set myvalue = 'x', lastupdated = UPDATE_COMMENCED
where id in
(
select id from foo where lastupdated < UPDATE_COMMENCED
limit SOME_REASONABLE_NUMBER
)
P.S. With respect to slowness:
I also tried UPDATE with LIMIT and also WHERE custom_condition(C1). These approaches do allow me to terminate update earlier, but they are significantly slower than regular update...
If you're willing to give other processes access to stale data, and your update is designed so as not to hog system resources, why is there a need to have the update complete within a certain amount of time? There seems to be no need to worry about perfomance in absolute terms. The concern should be relative to other processes -- make sure they're not blocked.

Answer 3

I also posted this question at
http://thread.gmane.org/gmane.comp.db.sqlite.general/81946
and got several interesting answers, such as:
divide range of rowid into slices and update one slice at a time
use AUTOINCREMENT feature to start new update at the place where the previous update ended (by LIMIT 10000)
create a trigger that calls select raise(fail, ...) to abort update without rollback