Re: Zedstore - compressed in-core columnar storage - Mailing list pgsql-hackers
| From | Robert Eckhardt |
|---|---|
| Subject | Re: Zedstore - compressed in-core columnar storage |
| Date | |
| Msg-id | 97B660BE-7336-4966-8290-0F5E3CA1AD79@vmware.com Whole thread Raw |
| In response to | Re: Zedstore - compressed in-core columnar storage (Simon Riggs <simon@2ndquadrant.com>) |
| Responses |
Re: Zedstore - compressed in-core columnar storage
|
| List | pgsql-hackers |
On Dec 31, 2020, at 9:22 AM, Simon Riggs <simon@2ndquadrant.com> wrote:On Wed, 18 Nov 2020 at 00:31, Jacob Champion <pchampion@vmware.com> wrote:
So that's in need of resolution. I'd expect gin and gist to be pretty
flaky until we fix that.
Jacob and Soumyadeep,
Thanks for submitting this. I think a fix is still outstanding? and
the patch fails to apply on HEAD in two places.
Please can you submit the next version?
Do you mind if we add this for review to the Jan CF?
It is a lot of code and I think there is significant difficulty for
the community to accept that as-is, even though it looks to be a very
high quality submission. So I would like to suggest a strategy for
commit: we accept Zedstore as "Beta" or "Experimental" in PG14,
perhaps with a WARNING/Caution similar to the one that used to be
given by Postgres in earlier versions when you created a Hash index.
We keep Zedstore in "Beta" mode until a later release, PG15 or later
when we can declare Zedstore fully safe. That approach allows us to
get this into the repo asap, and then be fixed and improved
incrementally from here.
The goal for Zedstore is to get a Column Store into Postgres, but not necessarily Zedstore. (Zedstore itself would be nice) When designing Zedstore success for us would be:
- significantly more performant on OLAP type queries,
- performant enough to not be terrible with OLTP type queries
- must support compression
- cannot be append only, this was the case initially with Greenplum Column Store and it was a mistake. Customers want to update and delete
- it needs to be feature complete as compared to HEAP unless it doesn’t make sense
Our initial goal is to get the TableAM and executor molded into a state where the above is possible for anyone wanting a column store implementation.
Given the goal of addressing API/Executor issues generically first, we have been trying to peel off and work on the parts that are not tightly linked to Zedstore. Specifically I don’t think it would be ok to merge Zedstore into core when it might affect the performance of HEAP relations.
Instead of focusing on the larger, more difficult to review Zedstore patch, we are trying to peel off the touch points where Zedstore and the current server interact. Note this isn’t intended to be an exhaustive list, rather a list of the most immediate issues. Some of these issues are critical for Zedstore to work, i.e. column projection, while some of these issues point more towards ensuring the various layers in the code are clean so that folks leveraging the TableAM don’t need to write their own bits from whole cloth but rather can leverage appropriately generic primitives, i.e. DBsize or page inspect.
As such, an incomplete list of things currently on our radar:
1) Column Projection — We have a patch [1] that is a demonstration of what we would like to do. There are several TODOs in the email that can/will be addressed if the general method is acceptable
2) DBSize —Georgios has a patch [2] that begins to make DBSize less HEAP specific
3) Reloptions —Jeff Davis has a patch [3] that begins to make these more flexible, having spoken with him we think additional work needs to be done here
4) PageInspect —needs to be less HEAP specific but no work has been done here that I’m aware of
5) bitmapHeapScan —currently scans both the index and the relation, there are code comments to address this and we need to look into what a fix would mean
6) Bulk insertion —Justin Pryzby has a patch [4] we are following along with.
7) analyze — Denis has a patch which starts to address this [5]
Ideally we can peel out anything that is useful to any column store. Once those have been discussed and committed the general code should be in better shape as well.
— Rob
e.g.
"NOTICE: Caution: Zedstore is an experimental feature in PostgreSQL14
intended for robustness and performance testing only. Your data and/or
query accuracy may be at risk if you rely on this."
--
Simon Riggs http://www.EnterpriseDB.com/
pgsql-hackers by date: