|From:||Andres Freund <andres(at)anarazel(dot)de>|
|To:||Ashwin Agrawal <aagrawal(at)pivotal(dot)io>|
|Cc:||PostgreSQL mailing lists <pgsql-hackers(at)postgresql(dot)org>|
|Subject:||Re: Zedstore - compressed in-core columnar storage|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
On 2019-04-08 17:27:05 -0700, Ashwin Agrawal wrote:
> Heikki and I have been hacking recently for few weeks to implement
> in-core columnar storage for PostgreSQL. Here's the design and initial
> implementation of Zedstore, compressed in-core columnar storage (table
> access method).
That's very cool.
> Motivations / Objectives
> * Performance improvement for queries selecting subset of columns
> (reduced IO).
> * Reduced on-disk footprint compared to heap table. Shorter tuple
> headers and also leveraging compression of similar type data
> * Be first-class citizen in the Postgres architecture (tables data can
> just independently live in columnar storage)
> * Fully MVCC compliant
> * All Indexes supported
> * Hybrid row-column store, where some columns are stored together, and
> others separately. Provide flexibility of granularity on how to
> divide the columns. Columns accessed together can be stored
> * Provide better control over bloat (similar to zheap)
> * Eliminate need for separate toast tables
> * Faster add / drop column or changing data type of column by avoiding
> full rewrite of the table.
Is storage going through the bufmgr.c or separately?
> In uncompressed form, the page can be arbitrarily large. But after
> compression, it must fit into a physical 8k block. If on insert or
> update of a tuple, the page cannot be compressed below 8k anymore, the
> page is split. Note that because TIDs are logical rather than physical
> identifiers, we can freely move tuples from one physical page to
> another during page split. A tuple's TID never changes.
When does compression happen? After each modifcation of the expanded
"page"? Are repeated expansions prevented somehow, e.g. if I
insert/delete rows into the same page one-by-one?
> A metapage at block 0, has links to the roots of the B-trees. Leaf
> pages look the same, but instead of storing the whole tuple, stores
> just a single attribute. To reconstruct a row with given TID, scan
> descends down the B-trees for all the columns using that TID, and
> fetches all attributes. Likewise, a sequential scan walks all the
> B-trees in lockstep.
Does the size of the metapage limit the number of column [groups]? Or is
there some overflow / tree of trees / whatnot happening?
> Inserting a new row, splits the row into datums. Then for first column
> decide which block to insert the same to, and pick a TID for it, and
> write undo record for the same. Rest of the columns are inserted using
> that same TID and point to same undo position.
Is there some buffering? Without that it seems like retail inserts are
going to be pretty slow?
> Property is added to Table AM to convey if column projection is
> leveraged by AM for scans. While scanning tables with AM leveraging
> this property, executor parses the plan. Leverages the target list and
> quals to find the required columns for query. This list is passed down
> to AM on beginscan. Zedstore uses this column projection list to only
> pull data from selected columns. Virtual tuple table slot is used to
> pass back the datums for subset of columns.
> Current table am API requires enhancement here to pass down column
> projection to AM. The patch showcases two different ways for the same.
> * For sequential scans added new beginscan_with_column_projection()
> API. Executor checks AM property and if it leverages column
> projection uses this new API else normal beginscan() API.
> * For index scans instead of modifying the begin scan API, added new
> API to specifically pass column projection list after calling begin
> scan to populate the scan descriptor but before fetching the tuples.
FWIW, I don't quite think this is the right approach. I've only a vague
sketch of this in my head, but I think we should want a general API to
pass that down to *any* scan. Even for heap, not deforming leading
columns that a uninteresting, but precede relevant columns, would be
quite a noticable performance win. I don't think the projection list is
the right approach for that.
> Extremely basic UNDO logging has be implemented just for MVCC
> perspective. MVCC is missing tuple lock right now. Plus, doesn't
> actually perform any undo yet. No WAL logging exist currently hence
> its not crash safe either.
Have you looked at the undo APIs developed for zheap, as discussed on
the list? Seems important that they're suitable for this too.
> Test mentioned in thread "Column lookup in a row performance" ,
> good example query for zedstore locally on laptop using lz4 shows
> postgres=# SELECT AVG(i199) FROM (select i199 from layout offset 0) x; --
> (1 row)
> Time: 4679.026 ms (00:04.679)
> postgres=# SELECT AVG(i199) FROM (select i199 from zlayout offset 0) x; --
> (1 row)
> Time: 379.710 ms
Well, I'm not sure I'm actually impressed by that. What does the
performance look like if you select i0 instead?
> Important note:
> Planner has not been modified yet to leverage the columnar
> storage. Hence, plans using "physical tlist" optimization or such good
> for row store miss out to leverage the columnar nature
> currently. Hence, can see the need for subquery with OFFSET 0 above to
> disable the optimization and scan only required column.
I'm more and more thinking that we should just nix the physical tlist
stuff and start afresh.
Congrats again, this is cool stuff.
|Next Message||Kyotaro HORIGUCHI||2019-04-09 01:28:48||Re: Problem with default partition pruning|
|Previous Message||Amit Langote||2019-04-09 01:00:41||Re: [Sender Address Forgery]Re: [HACKERS] generated columns|