Re: [PATCH 2/4] vinyl: do not use index lsn to identify indexes in vylog

[Vladimir Davydov <vdavydov.dev@gmail.com>]

Please ignore this patch for now. We can't use space_id/index_id for
identifying indexes in vylog after all. The problem is during ALTER we
have to maintain two indexes with the same space_id/index_id. It seems
that we need a unique index identifier in vylog after all, and it can't
be LSN. I'll rework the patch.

On Sat, Mar 17, 2018 at 08:56:35PM +0300, Vladimir Davydov wrote:
> An index can be dropped and then recreated with the same space/index id.
> To discriminate between different incarnations of the same index during
> recovery, we use LSN from the time of index creation, as it is supposed
> to be unique. However, the uniqueness property won't hold once ALTER is
> implemented for vinyl spaces: the problem is we have to rebuild all
> indexes of a space and commit them simultaneously in case the primary
> index is ALTERED.
> 
> Actually, we don't really need a unique index ID stored both in vylog
> and in WAL to recover all incarnations of vinyl indexes. Since WAL is
> replayed in the chronological order, we can find the index corresponding
> to index creation statement replayed from WAL by its space_id/index_id
> and the number of times the index has been dropped since we started to
> replay WAL. As we don't actually recover indexes that are dropped during
> WAL recovery, we don't need to know their properties, we just need to
> know that they existed. So we don't need to store a separate object for
> each index incarnation in the recovery context, it's enough to add a
> counter representing the number of incarnations the index has had since
> the last checkpoint to the vy_index_recovery_info structure to be able
> to recover all incarnations of the same vinyl index. Then on recovery,
> vy_index_recover() would check the incarnation counter of the index and
> 
>  - recover the last incarnation if it is 1
>  - emit create/drop record if it is > 1
>  - do nothing if it is 0
> 
> and decrement the counter on success so that the next call to
> vy_index_recover() with the same space and index id proceeds to
> the next incarnation of this index. And this is what this patch
> does in a nutshell.
> 
> The tricky part is preserving backward compatibility. We can't just
> replace index_lsn with space_id/index_id everywhere, because there
> still can be records that use index_lsn to identify an index and we
> need to process them. For example, VY_LOG_CREATE_INDEX is fine in
> this regard, because it contains both space_id/index_id and index_lsn,
> but VY_LOG_DROP_INDEX isn't, as it only has index_lsn. So we have to
> maintain two index hash tables during vylog recovery instead of just
> one: one for new records where index is referenced by space_id/index_id
> and another for legacy records where index_lsn is used instead.
> ---
>  src/box/vinyl.c          |  39 ++-----
>  src/box/vy_index.c       |  54 ++++-----
>  src/box/vy_index.h       |  12 +-
>  src/box/vy_log.c         | 286 ++++++++++++++++++++++++++++-------------------
>  src/box/vy_log.h         |  97 +++++++++++-----
>  src/box/vy_scheduler.c   |  12 +-
>  test/unit/vy_log_stub.c  |   2 +-
>  test/vinyl/layout.result |  32 +++---
>  8 files changed, 300 insertions(+), 234 deletions(-)