[openstack-dev] [oslo.db] Marker based paging

Heald, Mike mike.heald at hp.com
Thu Nov 6 09:47:49 UTC 2014

Unfortunately, ordering on anything but the timestamp
results in problems with usability and results consistency.

Example: Ordering on priority so I can see all my high priority tasks
first, then medium, then low.

Works ok, as long as the underlying data is geared so that the priorities
are ordered that way. But I have no way of skipping pages to find
them if they're not, and if the task that that we're using as the marker
changes priority as I page, I'm jumped to somewhere else in the list and I
have no way of knowing where I am in the list. I'd either see a load of
repeated data (with no way of skipping it)  or miss out on a lot of data
with no indication that this has happened.

So that's quite specific to Storyboard, but think of a situation where
you've ordered your instances in horizon by status so you can see
all that are spawning, running, deleting, etc. grouped together.
As you page, the marker's instance goes from spawning to running,
or running to deleting.

These aren't edge cases. The nature of the data we deal with forces
us to handle results that are going to change beneath us.

Now, I'm in no way suggesting that offset/limit would be a better solution!
But we do need to document these use cases somewhere and make sure
our paging solution can fulfil them, because Storyboard cannot provide the
user experience it needs to with the current pagination method in oslo.db.


From: Jay Pipes [jaypipes at gmail.com]
Sent: 04 November 2014 07:07
To: openstack-dev at lists.openstack.org
Subject: Re: [openstack-dev] [oslo.db] Marker based paging

On 11/04/2014 01:08 AM, Heald, Mike wrote:
> Thanks for that, Steven :)
> So just to clarify, results are ordered by the relevant timestamps to
> ensure consistent order and so that new records would never show on
> previous "pages" and be missed, and we're limited to just a "next
> page" navigation, and we cannot order the entire result set on any
> column but the timestamps, as this would break the paging because we
> can't do the comparisons we need to if the results aren't in that
> order. Have I got that correct?

No, that's not correct. There's nothing limiting one from ordering on
other columns than timestamp. We always ensure that there is a secondary
order on a column with unique values (like the primary key), in order to
ensure that pages of results are strictly ordered even when the sort
field is non-unique (like timestamp).

We're limited to next-previous pagination by choice because of the
scalability and performance limitations of a limit-offset pagination


OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org

More information about the OpenStack-dev mailing list