Обсуждение: Why does CREATE INDEX CONCURRENTLY need two scans?

On Wed, Apr 1, 2015 at 9:43 AM, Joshua Ma <josh@benchling.com> wrote:

Hi all,

I was curious about why CONCURRENTLY needs two scans to complete - from the documentation on HOT (access/heap/README.HOT), it looks like the process is:

1) insert pg_index entry, wait for relevant in-progress txns to finish (before marking index open for inserts, so HOT updates won't write incorrect index entries)

2) build index in 1st snapshot, mark index open for inserts

3) in 2nd snapshot, validate index and insert missing tuples since first snapshot, mark index valid for searches

Why are two scans necessary? What would break if it did something like the following?

1) insert pg_index entry, wait for relevant txns to finish, mark index open for inserts 

2) build index in a single snapshot, mark index valid for searches

Wouldn't new inserts update the index correctly? Between the snapshot and index-updating txns afterwards, wouldn't all updates be covered?

When an index is built with index_build, are included in the index only the tuples seen at the start of the first scan. A second scan is needed to add in the index entries for the tuples that have been inserted into the table during the build phase.
--

Michael

Michael Paquier <michael.paquier@gmail.com> writes:
> On Wed, Apr 1, 2015 at 9:43 AM, Joshua Ma <josh@benchling.com> wrote:
>> Why are two scans necessary? What would break if it did something like the
>> following?
>>
>> 1) insert pg_index entry, wait for relevant txns to finish, mark index
>> open for inserts
>>
>> 2) build index in a single snapshot, mark index valid for searches

>> Wouldn't new inserts update the index correctly? Between the snapshot and
>> index-updating txns afterwards, wouldn't all updates be covered?

> When an index is built with index_build, are included in the index only the
> tuples seen at the start of the first scan. A second scan is needed to add
> in the index entries for the tuples that have been inserted into the table
> during the build phase.

More to the point: Joshua's design supposes that retail insertions into
an index can happen in parallel with index build.  Or in other words,
that index build consists of instantaneously creating an empty-but-valid
index file and then doing a lot of ordinary inserts into it.  That's a
possible design, but it's not very efficient, and most of our index AMs
don't do it that way.  btree, for instance, starts by sorting all the
entries and creating the leaf-level pages.  Then it builds the upper tree
levels.  It doesn't have a complete tree that could support retail
insertions until the very end.  Moreover, most of the work is done in
storage that's local to the backend running CREATE INDEX, and isn't
accessible to other processes at all.

                        regards, tom lane