|From:||Yugo Nagata <nagata(at)sraoss(dot)co(dot)jp>|
|Cc:||Kevin Grittner <kgrittn(at)gmail(dot)com>|
|Subject:||Implementing Incremental View Maintenance|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
I would like to implement Incremental View Maintenance (IVM) on PostgreSQL.
IVM is a technique to maintain materialized views which computes and applies
only the incremental changes to the materialized views rather than
recomputate the contents as the current REFRESH command does.
I had a presentation on our PoC implementation of IVM at PGConf.eu 2018 .
Our implementation uses row OIDs to compute deltas for materialized views.
The basic idea is that if we have information about which rows in base tables
are contributing to generate a certain row in a matview then we can identify
the affected rows when a base table is updated. This is based on an idea of
Dr. Masunaga  who is a member of our group and inspired from ID-based
In our implementation, the mapping of the row OIDs of the materialized view
and the base tables are stored in "OID map". When a base relation is modified,
AFTER trigger is executed and the delta is recorded in delta tables using
the transition table feature. The accual udpate of the matview is triggerd
by REFRESH command with INCREMENTALLY option.
However, we realize problems of our implementation. First, WITH OIDS will
be removed since PG12, so OIDs are no longer available. Besides this, it would
be hard to implement this since it needs many changes of executor nodes to
collect base tables's OIDs during execuing a query. Also, the cost of maintaining
OID map would be high.
For these reasons, we started to think to implement IVM without relying on OIDs
and made a bit more surveys.
We also looked at Kevin Grittner's discussion  on incremental matview
maintenance. In this discussion, Kevin proposed to use counting algorithm 
to handle projection views (using DISTNICT) properly. This algorithm need an
additional system column, count_t, in materialized views and delta tables of
However, the discussion about IVM is now stoped, so we would like to restart and
Through our PoC inplementation and surveys, I think we need to think at least
the followings for implementing IVM.
1. How to extract changes on base tables
I think there would be at least two approaches for it.
- Using transition table in AFTER triggers
- Extracting changes from WAL using logical decoding
In our PoC implementation, we used AFTER trigger and transition tables, but using
logical decoding might be better from the point of performance of base table
If we can represent a change of UPDATE on a base table as query-like rather than
OLD and NEW, it may be possible to update the materialized view directly instead
of performing delete & insert.
2. How to compute the delta to be applied to materialized views
Essentially, IVM is based on relational algebra. Theorically, changes on base
tables are represented as deltas on this, like "R <- R + dR", and the delta on
the materialized view is computed using base table deltas based on "change
propagation equations". For implementation, we have to derive the equation from
the view definition query (Query tree, or Plan tree?) and describe this as SQL
query to compulte delta to be applied to the materialized view.
There could be several operations for view definition: selection, projection,
join, aggregation, union, difference, intersection, etc. If we can prepare a
module for each operation, it makes IVM extensable, so we can start a simple
view definition, and then support more complex views.
3. How to identify rows to be modifed in materialized views
When applying the delta to the materialized view, we have to identify which row
in the matview is corresponding to a row in the delta. A naive method is matching
by using all columns in a tuple, but clearly this is unefficient. If thematerialized
view has unique index, we can use this. Maybe, we have to force materialized views
to have all primary key colums in their base tables. In our PoC implementation, we
used OID to identify rows, but this will be no longer available as said above.
4. When to maintain materialized views
There are two candidates of the timing of maintenance, immediate (eager) or deferred.
In eager maintenance, the materialized view is updated in the same transaction
where the base table is updated. In deferred maintenance, this is done after the
transaction is commited, for example, when view is accessed, as a response to user
In the previous discussion, it is planned to start from "eager" approach. In our PoC
implementaion, we used the other aproach, that is, using REFRESH command to perform IVM.
I am not sure which is better as a start point, but I begin to think that the eager
approach may be more simple since we don't have to maintain base table changes in other
In the eager maintenance approache, we have to consider a race condition where two
different transactions change base tables simultaneously as discussed in .
 https://ipsj.ixsq.nii.ac.jp/ej/index.php?active_action=repository_view_main_item_detail&page_id=13&block_id=8&item_id=191254&item_no=1 (Japanese only)
Yugo Nagata <nagata(at)sraoss(dot)co(dot)jp>
|Next Message||Alvaro Herrera||2018-12-27 13:15:08||Re: Feature: temporary materialized views|
|Previous Message||Magnus Hagander||2018-12-27 12:54:34||Re: global / super barriers (for checksums)|