[Tarantool-patches] [PATCH 2/2] vinyl: skip invalid upserts during squash

[Nikita Pettik]

On 19 May 22:39, Konstantin Osipov wrote:
> * Nikita Pettik <korablev at tarantool.org> [20/05/19 22:10]:
> > On 14 May 09:56, Konstantin Osipov wrote:
> > > * Nikita Pettik <korablev at tarantool.org> [20/05/14 09:50]:
> > > > On 14 Apr 01:12, Konstantin Osipov wrote:
> > > > > * Nikita Pettik <korablev at tarantool.org> [20/04/14 00:57]:
> > > 
> > > > > Finally, let's double check there are no issues with the used
> > > > > format - can it become obsolete by the time it's used, e.g. if
> > > > > there is an online/non-blocking schema change that happened in tx
> > > > > thread (compaction is running in the write thread)?
> > > > 
> > > > Upserts are supported only by primary index. Meanwhile vinyl does
> > > > not support altering primary index of non-empty space. Am I missing
> > > > something?
> > > 
> > > I mean the space format object itself. Squashing is happening in a
> > > different thread, the 
> > > > 
> > 
> > Seems like you've cut sentence. Could you please suggest exact
> > test scenario? We already have similar test in vinyl/errinj.test.lua:
> 
> 1. Memtable dump is started.
> 2. Space is altered.
> 3. The format object is gone.
> 4. Memtable dump uses the old format object for upsert
>    squashing.

Look, the format is used for upsert squashing is stored in write
iterator: in vy_read_view_merge() it is passed to vy_apply_upsert()
from stream->format. When write iterator is created, format which
is obtained (i.e. lsm->disk_format) is refed. So, as you can see,
it can't be destroyed before write iterator is finished.

Finally, as I already mentioned, lsm used for squashing is PK,
and PK in turn can't be altered (lsm->disk_format will be destroyed
only on pk:drop()).

> Similar with compaction:
> 1. Compaction starts and uses the existing format objects
> 2. It uses the format object for upsert squashing
> 3. Space is altered and format is changed
> 4. We report a squash failure/suppress failure.
> 
> 
> This is both related to memory liveness issues (e.g. we may use a
> freed format object, which isn't the case because we refcount them
> at start of dump/compaction), but also correctness issues: we
> accept data which is no longer acceptable according to the new
> format, or report/skip data which shouldn't be skipped.

Data will always will be acceptable according to the format because
PK format can't change.

> -- 
> Konstantin Osipov, Moscow, Russia