22.05.2018 01:03, Eric Blake wrote:
On 05/21/2018 10:52 AM, Vladimir Sementsov-Ogievskiy wrote:
18.05.2018 01:43, Eric Blake wrote:
Here's my updated counterproposal for a backup API.
[...]
Representing things on a timeline, when a guest is first created,
there is no dirty bitmap; later, the checkpoint "check1" is created,
which in turn creates "bitmap1" in the qcow2 image for all changes
past that point; when a second checkmark "check2" is created, a qemu
transaction is used to create and enable the new "bitmap2" bitmap at
the same time as disabling "bitmap1" bitmap. (Actually, it's probably
easier to name the bitmap in the qcow2 file with the same name as the
Checkpoint object being tracked in libvirt, but for discussion
purposes, it's less confusing if I use separate names for now.)
creation ....... check1 ....... check2 ....... active
no bitmap bitmap1 bitmap2
When a user wants to create a backup, they select which point in time
the backup starts from; the default value NULL represents a full
backup (all content since disk creation to the point in time of the
backup call, no bitmap is needed, use sync=full for push model or
sync=none for the pull model); any other value represents the name of
a checkpoint to use as an incremental backup (all content from the
checkpoint to the point in time of the backup call; libvirt forms a
temporary bitmap as needed, the uses sync=incremental for push model
or sync=none plus exporting the bitmap for the pull model). For
example, requesting an incremental backup from "check2" can just reuse
"bitmap2", but requesting an incremental backup from "check1" requires
the computation of the bitmap containing the union of "bitmap1" and
"bitmap2".
I have a bit of criticism on this part, exactly on ability to create
a backup not from last checkpoint but from any from the past. For
this ability we are implementing the whole api with checkpoints, we
are going to store several bitmaps in Qemu (and possibly going to
implement checkpoints in Qemu in future). But personally, I don't
know any real and adequate use cases for this ability.
I heard about the following cases:
1. Incremental restore: we want to rollback to some point in time
(some element in incremental backup chain), and don't want to copy
all the data, but only changed.
- It's not real case, because information about dirtiness is already
in backup chain: we just need to find allocated areas and copy them +
we should copy areas, corresponding to dirty bits in active dirty
bitmap in Qemu.
If you do a pull mode backup (where the dirty bitmaps were exported
over NBD), then yes, you can assume that the third-party app reading
the backup data also saved the dirty bitmap in whatever form it likes,
so that it only ever has to pull data from the most recent checkpoint
and can reconstruct the union of changes from an earlier checkpoint
offline without qemu help. But for a push mode backup (where qemu
does the pushing), there is no way to expose the dirty bitmap of what
the backup contains, unless you backup to something like a qcow2 image
and track which clusters in the backup image were allocated as a
result of the backup operation. So having a way in the libvirt API to
grab an incremental backup from earlier than the most recent
checkpoint may not be needed by everyone, but I don't see a problem in
implementing it either.
If we have a chain of incrementals from push backup, it should be
possible to analyze their block status, so it should be something like
qcow2. Otherwise we can't create a chain of backing files. If we backup
incrementals to the same file, we anyway can't restore to some previous
point except the last one.
2. Several backup solutions backing up the same vm
- Ok, if we implement checkpoints, instead of maintaining several
active dirty bitmaps, we can have only one active bitmap and others
disabled, which lead to performance gain and possibility to save RAM
space (if we unload disabled bitmaps from RAM to qcow2). But what are
real cases? What is the real benefit? I doubt that somebody will use
more than 2 - 3 different backup providers on same vm, so is it worth
implementing such a big feature for this? It of course worth doing if
we have 100 independent backup providers.
Note: the word "independent" is important here. For example it may be
two external backup tools, managed by different subsystems or
different people or something like this. If we are just doing a
backup weekly + daily, actually, we can synchronize them, so that
weekly backup will be a merge of last 7 daily backups, so weekly
backup don't need personal active dirty bitmap and even backup
operation.
I'm not sure if this is a complaint that libvirt should allow more
than one active bitmap at a time, vs. having exactly one active bitmap
at a time and then reconstructing bitmaps over larger sequences of
time as needed. But does that change the API that libvirt should
expose to end users, or can it just be an implementation detail?
[2] is a benefit from new API, which allows such implementation detail
(only one active bitmaps). And I just say, that this benefit looks not
very significant.
3. Some of backups in incremental backup chain are lost, and we want
to recreate part of the chain as a new backup, instead of just
dropping all chain and create full backup.
In this case, I can say the following:
disabled bitmaps (~ all checkpoints except the last one) are constant
metadata, related to the backup chain, not to the vm. And it should
be stored as constant data: may be on the same server as backup
chain, maybe on the other, maybe in some database, but not in vm. VM
is a dynamic structure, and I don't see any reason of storing
(almost) unrelated constant metadata in it. Also, saving this
constant backup-related metadata separately from vm will allow to
check it's consistency with a help of checksums or something like
this. Finally, I'm not a specialist in storing constant data, but I
think that the vm is not the best place.
How does backup bitmap data get lost? If the image is only managed by
libvirt, then libvirt shouldn't be losing arbitrary bitmaps. If a
third-part entity is modifying the qcow2 images (presumably while the
guest is offline, as editing a file that is simultaneously in use by
qemu is a no-no), then all bets are off anyways, as you really
shouldn't be trying to independently manage qcow2 files that are
already being managed by libvirt.
Note: Hmm, do someone have real examples of such user cases? Why
backups are lost, is it often case? (I heard an assumption, that it
may be a tool, checking backups (for example create a vm over the
backup and check that it at least can start), which is running in
background. But I'm not sure, that we must drop backup if it failed,
may be it's enough to merge it up)
I'm more worried about the implementation of checkpoints and dirty
bitmaps across domain snapshots (where we have to consider copying one
or more bitmaps from the base image to the active image when creating
a snapshot, and conversely about merging bitmaps from the active image
into the base when doing a live commit).
Hm, I think we don't need merge, but copy in both operations. Both
operations don't change the data from guest point of view, so bitmap
should not change, just copied to the topmost image. And anyway on
commit, we should drop all old bitmaps from base image (or update them
by dirtying bits, corresponding to allocated clusters in top image).
One related note here: we also can use dirty-bitmaps migration
capability to save bitmaps to vmstate.
3.1 About external backup: we have even already exported this
metadata to the third backup tool. So, this tool should store this
information for future use, instead of exporting from Qemu again.
To summarize:
1. I doubt that discussed ability is really needed.
2. If it is needed, I doubt that it's a true way to store related
disabled bitmaps (or checkpoints) in Qemu.
So, to make sure I understand, the only thing that you are debating
whether we need is the ability to grab a backup image from earlier
than the most-recent checkpoint? Remember, the proposal is whether we
have a sufficiently powerful libvirt API to cover multiple use cases,
even if not all users need all of the permutations of use cases, while
still being something that is concise enough to document and implement
on top of existing qemu semantics.
I agree, that checkpoint-based API is powerful. I just don't see real
benefits (use cases).
--
Best regards,
Vladimir
--
libvir-list mailing list
libvir-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/libvir-list
