|
|
Posted by sql_server_user on 05/11/06 01:17
Thanks, Erland - I'm gonna have to pay you a consulting fee soon.
The recovery model is set to simple, there are no nonclustered indexes
on the table, and I agree that clustering on insert is the way to go,
since I think trying to recluster the new table with all the data in it
would just blow up the logs and tempdb like my initial insert attempt.
I like the idea of the batched insert with the low fill factor
clustered index, however, I ran a bunch of tests and the amount of log
and tempdb space required seems to be a function of the size of the
destination table, not the number of rows in the batch: I would insert
a batch (10 million rows), shrink the log and tempdb, insert another
batch, and so on, and every time, the log and tempdb would require more
space, roughly linearly as my very approximate data below indicate:
after 1st batch: log = 1GB, tempdb = 2GB
after shrinking and 2nd batch: log = 2GB, tempdb = 4GB
after shrinking and 3rd batch: log = 3 GB, tempdb = 6GB
So I'm worried about the batch approach for the full insert.
I'm currently trying an SSIS copy from one table to the other, and
since it's doing a bulk insert it doesn't seem to be touching the db
log. However, tempdb is getting pretty fat, 12GB and going strong, but
I've cleared a lot of room for it this time - it has about 179GB to go,
so hopefully that will be enough. If this doesn't work I'll definitely
try your batched insert suggestion.
Thanks again,
Seth
Erland Sommarskog wrote:
> sql_server_user (kaioptera@gmail.com) writes:
> > I'm trying to copy all 440 million rows from one table in my SQL Server
> > 2005 db to another table with a different clustering scheme. After a
> > few test inserts that were successful (up to a million rows in 20
> > seconds with no problem), I crossed my fingers and went whole-hog with
> > the naive "insert into dest_table (column1, column2, ...) select
> > column1, column2, ... from source_table" but of course it ran out of
> > space in *both* the db log file (at about 130 GB) and the tempdb log
> > file (at about 54GB) and rolled it back, wrapping up 10 hours after I
> > kicked it off.
>
> That operation is definitely not a walk in a park.
>
> I think I would attempt this strategy:
>
> 1) Set recovery for the database to simple recovery.
> 2) Create the clustered index on the target table with a relatively low
> fill factor, say 30%. (Of course, no NC indexes would be in place.)
> 3) Insert batchwise from the old table, where batches are defined by
> intervals of the clustered index of the source table. For batchsize,
> I would use 100000 to 500000 rows, depending on wide the table is.
> 4) I would make the control loop so that it is easy to stop it. For
> instance on each iteration read from a control table. Possibly also
> a table the defines the batches, so you can keep track of what has
> been copied. But the table should not be joined with the INSERT;
> read interval into variales.
> 5) Occassionally stop the process and monitor fragmentation in
> target table. It should first decrease, as holes are filled in,
> but then it will increase again. (But you will probably see this
> from the fact that the time for iteration increases.)
> 6) Once everything is done, switch to full recovery and backup the
> database.
>
> My thinking here is that creating the clustered index on this monster
> is going require a lot of log and tempdb - you can guess what happens.
> So let's have the clustered index in place from the start, even if
> that will take longer time.
>
>
>
> --
> Erland Sommarskog, SQL Server MVP, esquel@sommarskog.se
>
> Books Online for SQL Server 2005 at
> http://www.microsoft.com/technet/prodtechnol/sql/2005/downloads/books.mspx
> Books Online for SQL Server 2000 at
> http://www.microsoft.com/sql/prodinfo/previousversions/books.mspx
Navigation:
[Reply to this message]
|