On Sun, Sep 17, 2000 at 09:42:44AM -0700, K Parker wrote:
> Sure you can lock a single row:
>
> BEGIN;
> SELECT .... FOR UPDATE ... ;
> -- app messes with filesystem...
> -- may or may not actually update the row
> COMMIT;
>
> Won't this do what you need?
Thanks - that's very useful to know.
However, it's not exactly what I'm looking for. I'll explain fully.
This table contains a list of messages. One of the fields is the
path to where the message is stored in plaintext on the filesystem
(it's not in the DB because it could be > 8K, I also don't want to
take up a lot of space). Another field is the boolean field 'archived'
- in other words, has this message been compressed - y/n?
So the app is doing regular maintainence. It goes through the messages,
and selects a list of the really old ones that are going to be compressed
to save space. Now, the app goes through the list, and does the actual
compression on disk. *Here* is where everything could get screwed
up - if someone queries the DB looking for a message that has been
compressed on the FS but the transaction hasn't yet been committed,
they'll read the compressed file off the HD and try to display it -
but it will be garbage. Of course, once the transaction is committed
and the 'archived' flag is toggled, the client app will know to
decompress the file before trying to read it. However, I'm worried
about the time in between the initial 'SELECT ... FOR UPDATE' and
when the transaction is committed.
One thing - the filename is the same whether the file is compressed or
not. So the client app can only look @ the 'archived' boolean to judge
if it needs to decompress the file.
If I wrap this inside a transaction, other queries will see the state of
the DB *before* the transaction was started. Which is bad - they will
see 'archived' is false, and read the compressed file from the FS as
though it was plaintext.
What I would like to happen is that if any queries want to access the row
while the transaction isn't committed, they should just poll until the
transaction is committed or rolled back. This way, they won't get *any*
results until the data is in a consistent state. Is this possible? Or am I
going about things the wrong way?
Thanks in advance,
Neil
--
Neil Conway <neilconway@home.com>
Get my GnuPG key from: http://klamath.dyndns.org/mykey.asc
Encrypted mail welcomed
Don't worry about people stealing your ideas. If your ideas are any
good, you'll have to ram them down people's throats.
-- Howard Aiken
