Re: Make drop database safer

On 2/12/19 12:55 AM, Ashwin Agrawal wrote:
> 
> Thanks for the response and inputs.
> 
> On Sat, Feb 9, 2019 at 4:51 AM Andres Freund <andres@anarazel.de 
> <mailto:andres@anarazel.de>> wrote:
> 
>     Hi,
> 
>     On 2019-02-08 16:36:13 -0800, Alexandra Wang wrote:
>      > Current sequence of operations for drop database (dropdb())
>      > 1. Start Transaction
>      > 2. Make catalog changes
>      > 3. Drop database buffers
>      > 4. Forget database fsync requests
>      > 5. Checkpoint
>      > 6. Delete database directory
>      > 7. Commit Transaction
>      >
>      > Problem
>      > This sequence is unsafe from couple of fronts. Like if drop database,
>      > aborts (means system can crash/shutdown can happen) right after
>     buffers are
>      > dropped step 3 or step 4. The database will still exist and fully
>      > accessible but will loose the data from the dirty buffers. This
>     seems very
>      > bad.
>      >
>      > Operation can abort after step 5 as well in which can the entries
>     remain in
>      > catalog but the database is not accessible. Which is bad as well
>     but not as
>      > severe as above case mentioned, where it exists but some stuff goes
>      > magically missing.
>      >
>      > Repo:
>      > ```
>      > CREATE DATABASE test;
>      > \c test
>      > CREATE TABLE t1(a int); CREATE TABLE t2(a int); CREATE TABLE t3(a
>     int);
>      > \c postgres
>      > DROP DATABASE test; <<====== kill the session after
>     DropDatabaseBuffers()
>      > (make sure to issue checkpoint before killing the session)
>      > ```
>      >
>      > Proposed ways to fix
>      > 1. CommitTransactionCommand() right after step 2. This makes it
>     fully safe
>      > as the catalog will have the database dropped. Files may still
>     exist on
>      > disk in some cases which is okay. This also makes it consistent
>     with the
>      > approach used in movedb().
> 
>     To me this seems bad. The current failure mode obviously isn't good, but
>     the data obviously isn't valuable, and just loosing track of an entire
>     database worth of data seems worse.
> 
> 
> So, based on that response seems not loosing track to the files 
> associated with the database is design choice we wish to achieve. Hence 
> catalog having entry but data directory being deleted is fine behavior 
> to have and doesn't need to be solved.
> 

What about adding 'is dropped' flag to pg_database, set it to true at 
the beginning of DROP DATABASE and commit? And ensure no one can connect 
to such database, making DROP DATABASE the only allowed operation?

ISTM we could then continue doing the same thing we do today, without 
any extra checkpoints etc.

>  > 2. Alternative way to make it safer is perform Checkpoint (step 5) just
> 
>      > before dropping database buffers, to avoid the unsafe nature.
>     Caveats of
>      > this solution is:
>      > - Performs IO for data which in success case anyways will get deleted
>      > - Still doesn't cover the case where catalog has the database
>     entry but
>      > files are removed from disk
> 
>     That seems like an unacceptable slowdown.
> 
> 
> Given dropping database should be infrequent operation and only addition 
> IO cost is for buffers for that database itself as Checkpoint is anyways 
> performed in later step, is it really unacceptable slowdown, compared to 
> safety it brings ?
> 

That's probably true, although I do know quite a few systems that create 
and drop databases fairly often. And the implied explicit checkpoints 
are quite painful, so I'd vote not to make this worse.

FWIW I don't recall why exactly we need the checkpoints, except perhaps 
to ensure the file copies see the most recent data (in CREATE DATABASE) 
and evict stuff for the to-be-dropped database from shared bufers. I 
wonder if we could do that without a checkpoint somehow ...

> 
>      > 3. One more fancier approach is to use pending delete mechanism
>     used by
>      > relation drops, to perform these non-catalog related activities
>     at commit.
>      > Easily, the pending delete structure can be added boolean to convey
>      > database directory dropping instead of file. Given drop database
>     can't be
>      > performed inside transaction, not needed to be done this way, but
>     this
>      > makes it one consistent approach used to deal with on-disk removal.
> 
>     ISTM we'd need to do something like this.
> 
> 
> Given the above design choice to retain link to database files till 
> actually deleted, not seeing why pending delete approach any better than 
> approach 1. This approach will allow us to track the database oid in 
> commit transaction xlog record but any checkpoint post the same still 
> looses the reference to the database. Which is same case in approach 1 
> where separate xlog record XLOG_DBASE_DROP is written just after 
> committing the transaction.
> When we proposed approach 3, we thought its functionally same as 
> approach 1 just differs in implementation. But your preference to this 
> approach and stating approach 1 as bad, reads as pending deletes 
> approach is functionally different, we would like to hear more how?
> 

Hmmm, I don't see how this is an improvement over option #1 either.

> Considering the design choice we must meet, seems approach 2, moving 
> Checkpoint from step 5 before step 3 would give us the safety desired 
> and retain the desired link to the database till we actually delete the 
> files for it.
> 

Ummm? That essentially means this order:

1. Start Transaction
2. Make catalog changes
5. Checkpoint
3. Drop database buffers
4. Forget database fsync requests
6. Delete database directory
7. Commit Transaction

I don't see how that actually fixes any of the issues? Can you explain?

Not to mention we might end up doing quite a bit of I/O to checkpoint 
buffers from the database that is going to disappear shortly ...

regards

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services