couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Robert Newson <rnew...@apache.org>
Subject Re: Even more fine-grained ETag support when querying views?
Date Tue, 13 Sep 2011 13:49:46 GMT
CRC32 should be good enough but there are better hash algorithms out
there (not completely sure they're commutative though). Will update.

On 13 September 2011 14:09, Paul Davis <paul.joseph.davis@gmail.com> wrote:
> Yeah, that's the basic idea. I walked through the idea of using
> something more familiar like SHA's or what not, but unless someone
> knows how to combine SHA hash states commutatively then I think that
> idea is shot because it'd cause a stampeding herd effect after
> compaction.
>
> On Tue, Sep 13, 2011 at 8:46 AM, Robert Newson <rnewson@apache.org> wrote:
>> My joke about bloom filters was apparently misunderstood but the
>> notion above, which sounds a lot like a Merkle tree, seems lucid to
>> me.
>>
>> As for the strong vs. weak ETag variants, I think views need strong
>> ETags in all cases, given the declared semantics for them in 13.3.3
>>
>> B.
>>
>> On 12 September 2011 23:28, Paul Davis <paul.joseph.davis@gmail.com> wrote:
>>> In general the idea is intriguing. Using a combining hash would allow
>>> you to get a specific hash value for a given range. Unfortunately,
>>> bloom filters are not a good solution here because they require an a
>>> priori guess of the number of keys that are going to be stored. On the
>>> other hand, CRC32 appears to be combinable.There are a couple issues
>>> though. The first of which is whether this is a strong enough hash to
>>> use for an ETag. There are two types of ETags with slightly different
>>> semantics, so we'd have to figure out what we can do and where this
>>> falls on that spectrum. Secondly, computing the range ETag would
>>> require the equivalent of a reduce=false view call in addition to
>>> streaming the output if validation matched which has performance
>>> implications as well.
>>>
>>> On Mon, Sep 12, 2011 at 6:50 PM, Alon Keren <alon.keren@gmail.com> wrote:
>>>> Disclosure: I don't know much about e-tags, CouchDB internals (or bloom
>>>> filters).
>>>>
>>>> How about maintaining an e-tag for each sub-tree in the view, similar to
the
>>>> way (I think) reduce works?
>>>> When a row gets updated, its e-tag would be recalculated, and then its
>>>> parent's e-tag would be recalculated, and so on. The e-tag of an internal
>>>> node could be the hash of all its children's hashes.
>>>> The actual e-tag that a view-query receives: the e-tag of the common
>>>> ancestor of all involved rows.
>>>>
>>>> The next time you query the same keys, you would supply the e-tag you've
>>>> just received.
>>>>
>>>>  Alon
>>>>
>>>>
>>>> On 10 September 2011 16:41, Andreas Lind Petersen <
>>>> andreaslindpetersen@gmail.com> wrote:
>>>>
>>>>> Hi!
>>>>>
>>>>> Background: I'm working on a web app that uses a single CouchDB database
>>>>> for
>>>>> storing data belong to 400000+ users. Each user has an average of about
40
>>>>> documents that need to be fetched in one go when the frontend is launched.
>>>>> I
>>>>> have accomplished this by querying a simple view with ?key=ownerID (with
a
>>>>> fallback to /_alldocs?startkey=<ownerID>_...&endkey=<ownerID>~
if the view
>>>>> isn't built). Since the data for each user rarely changes, there's a
>>>>> potential to save resources by supporting conditional GET with
>>>>> If-None-Match, which would amount having the web app backend copy the
>>>>> CouchDB-generated ETag into the response sent to the browser.
>>>>>
>>>>> However, I just learned that CouchDB only maintains a single ETag for
the
>>>>> entire view, so every time one of my users changes something, the ETag
for
>>>>> everyone else's query result also changes. This makes conditional GETs
>>>>> useless with this usage pattern.
>>>>>
>>>>> I asked about this on #couchdb and had a brief talk with rnewson, who
was
>>>>> sympathetic to the idea. Unfortunately we weren't able to come up with
an
>>>>> idea that didn't involve traversing all docs in the result just for
>>>>> computing the ETag (my suggestion was a hash of the _revs of all docs
>>>>> contributing to the result). That would be a bad default, but might still
>>>>> work as an opt-in thing per request, eg. slowetag=true.
>>>>>
>>>>> Newson said I should try raising the discussion here in case someone
else
>>>>> had an idea for a cheaper way to calculate a good ETag. So what does
>>>>> everyone else think about this? Is my use case too rare, or would it
be
>>>>> worthwhile to implement it?
>>>>>
>>>>> Best regards,
>>>>> Andreas Lind Petersen (papandreou)
>>>>>
>>>>> Here's our chat transcript:
>>>>>
>>>>> [11:46] <papandreou> Does anyone know if there are plans for issuing
even
>>>>> more granular etags for view lookups? When you only look up a small range
>>>>> or
>>>>> a specific key it would be really great if the ETag only changed when
that
>>>>> subset changes rather than the entire view.
>>>>> [11:47] <papandreou> In the application I'm working on I'll hardly
ever be
>>>>> able to get a 304 response because of this.
>>>>> [...]
>>>>> [13:51] <+rnewson> papandreou: unlikely.
>>>>> [13:52] <papandreou> rnewson: So the best thing I can do is to
fetch the
>>>>> data and compute a better etag myself? (My use case is a backend for
a web
>>>>> app)
>>>>> [13:53] <+rnewson> papandreou: You might be able to set ETag in
a list
>>>>> function? If you can't, I'll gladly change CouchDB so you can.
>>>>> [13:54] <papandreou> rnewson: I thought about that, too, but that
would
>>>>> cause a big overhead for every request, right?
>>>>> [13:55] <papandreou> rnewson: (Last time I tried views were slooow)
>>>>> [13:55] <papandreou> I mean lists
>>>>> [13:55] <+rnewson> papandreou: slower, yes, because couch needs
to evaluate
>>>>> the javascript in an external process.
>>>>> [13:55] <+rnewson> how will you calculate the fine-grained ETag?
>>>>> [13:56] <+rnewson> Also we did recently make it slightly finer,
before it
>>>>> was view group scope and now it's the view itself (I think)
>>>>> [13:56] <papandreou> rnewson: Maybe something like a hash of the
_revs of
>>>>> all the documents contributing to the result?
>>>>> [13:56] <+rnewson> hm, that makes no sense actually. but we did
refine it
>>>>> recently.
>>>>> [13:57] <+rnewson> papandreou: that doesn't sound cheap at all,
and it
>>>>> would
>>>>> need to be cheaper than doing the view query itself to make sense.
>>>>> [13:58] <papandreou> rnewson: There's still the bandwidth thing
>>>>> [13:58] <+rnewson> oh, you're working with restricted bandwidth
and/or have
>>>>> huge view responses?
>>>>> [13:59] <papandreou> rnewson: And it would be really nice to have
something
>>>>> like this completely handled by the database instead of inventing a bunch
>>>>> of
>>>>> workarounds.
>>>>> [14:01] <+rnewson> If there's a correct and efficient algorithm
for doing
>>>>> it, I'm sure it would be applied.
>>>>> [14:02] <papandreou> rnewson: I guess it depends on the use case.
If the
>>>>> database is rarely updated I suppose the current tradeoff is better.
>>>>> [14:03] <+rnewson> I'm sure the only reason we have ETags at the
current
>>>>> granularity is because it's very quick to calculate. A finer-grain would
be
>>>>> committed if a viable approach was proposed.
>>>>> [14:04] <papandreou> rnewson: I have a huge database with data
belonging to
>>>>> 400000+ different users, and I'm using a view to enable a lookup-by-owner
>>>>> thing. But every time a single piece of data is inserted, the ETag for
the
>>>>> view changes
>>>>> [14:04] == case_ [~case@AMontsouris-651-1-123-169.w83-202.abo.wanadoo.fr]
>>>>> has quit [Read error: Connection reset by peer]
>>>>> [14:04] <+rnewson> yes, I've completely understood the problem
you stated
>>>>> earlier.
>>>>> [14:05] <+rnewson> I can't think of a way to improve this right
now but I
>>>>> would spend the time to implement it if you had one.
>>>>> [14:06] <papandreou> rnewson: So right now the code path that sends
a 304
>>>>> only needs to look at a single piece of metadata for the view to make
its
>>>>> decision? That'll be hard to beat :)
>>>>> [14:07] <+rnewson> doesn't need to beat it, it just needs to be
fast.
>>>>> [14:07] <+rnewson> but I don't see any current possible solutions,
let
>>>>> alone
>>>>> fast ones.
>>>>> [14:07] <papandreou> rnewson: Well, thanks anyway for considering
my
>>>>> suggestion. I'll let you know of I get an idea :)
>>>>> [14:08] <+rnewson> and it is now per-view and not per-viewgroup.
so it's
>>>>> what I said first before I thought it was silly
>>>>> [14:08] <+benoitc> query + last seq returned maybe ....
>>>>> [14:08] <+rnewson> but obviously a change could affect one view
in a group
>>>>> but not others
>>>>> [14:09] <papandreou> benoitc: The query is already sort of included
since
>>>>> it's in the url.
>>>>> [14:09] <+rnewson> benoitc: ?
>>>>> [14:10] <+benoitc> i was meaning last committed seq,but it won't
change
>>>>> anything ...
>>>>> [14:10] <papandreou> benoitc: I guess you'd also need to make sure
that the
>>>>> ETag changes if a document is deleted?
>>>>> [14:10] <papandreou> ah
>>>>> [14:10] <+rnewson> benoitc: we already use the update_seq of the
#view,
>>>>> which is finer-grained that db's last committed seq
>>>>> [14:11] <+benoitc> rnewson: commited seq in the view group but
anyway it
>>>>> won't work
>>>>> [14:12] <+rnewson> benoitc: right, that would be the pre-1.1.0
behavior, I
>>>>> think.
>>>>> [14:12] <+rnewson> which is coarser
>>>>> [14:12] <+rnewson> we simply don't record the info that papandreou's
>>>>> suggestion would need to work.
>>>>> [14:12] <+benoitc> papandreou: easier solution would be to request
each
>>>>> time
>>>>> on on stale view
>>>>> [14:13] <papandreou> rnewson: Another reason why my suggestion
sucks is
>>>>> that
>>>>> it would require two traversals of the range, right? I'm guessing it
starts
>>>>> streaming as soon as it has found the first doc now?
>>>>> [14:13] <+benoitc> and update after, think it would work. except
if you
>>>>> want
>>>>> something strict
>>>>> [14:13] <+rnewson> papandreou: yes, we stream the results as we
read them,
>>>>> we don't buffer.
>>>>> [14:14] <papandreou> benoitc: Hmm, so the theory is that stale=ok
would
>>>>> increase the percentage of 304 responses?
>>>>> [14:14] <papandreou> rnewson: Right, yes, then it would take a
serious hit.
>>>>> [14:14] <+rnewson> papandreou: but we could add an option that
reads the
>>>>> thing, builds an etag, and then streams the result. it would be slower,
but
>>>>> for the times that we can send 304 we'd save bandwidth. It sounds a bit
too
>>>>> niche to me, but you could raise it on user@
>>>>> [14:15] == Frippe [~Frippe@unaffiliated/frippe] has quit [Ping timeout:
>>>>> 240
>>>>> seconds]
>>>>> [14:15] <papandreou> rnewson: Would be awesome to have that as
a
>>>>> configuration option
>>>>> [14:15] <+rnewson> papandreou: the view would not change, so neither
would
>>>>> the ETag (with stale=ok)
>>>>> [14:15] <+rnewson> papandreou: I think it would be a runtime option
>>>>> ?slow_etag=true
>>>>> [14:15] <papandreou> rnewson: That would also be fine
>>>>> [14:16] <+rnewson> a better solution would not require two passes,
though.
>>>>> [14:16] <+benoitc> papandreou: i would use stale=ok, then query
the view
>>>>> async, save new etag & ...
>>>>> [14:16] <papandreou> rnewson: I really don't think it's that niche
:). But
>>>>> maybe ETag-nerds are rarer than I think, hehe
>>>>> [14:16] <+benoitc> rnewson: that could encourage pretty dangerous
things
>>>>> [14:16] <+rnewson> benoitc: ?
>>>>> [14:17] <+benoitc> rnewson: cpu intensives tasks eacht time the
call is
>>>>> done,
>>>>> [14:17] <+benoitc> rather than encouraging something async
>>>>> [14:18] <+benoitc> rahh I hate osx, it introduce be bad unicode
chars in
>>>>> vim
>>>>> :@
>>>>> [14:23] == Frippe_ has changed nick to Frippe
>>>>> [14:23] <papandreou> benoitc: I'm not sure exactly how that would
work? I'm
>>>>> working on the backend for a web app, so the requests will be coming
from
>>>>> multiple machines
>>>>> [14:24] <+benoitc> papandreou: call with stale==ok and have a process
>>>>> asking
>>>>> your deb for refresh from time to time
>>>>> [14:24] <+benoitc> s/deb/view
>>>>> [14:25] <+rnewson> benoitc: not sure I follow. doubling the number
of view
>>>>> requests to achieve a finer etag is an ok solution, but shouldn't be
the
>>>>> default, but I do think we'd need a better solution than that.
>>>>> [14:25] <+rnewson> benoitc: and you might be forgetting all the
md5
>>>>> verification we do all the time.
>>>>> [14:27] <+benoitc> rnewson: you don't need to call each views though
>>>>> [14:27] <+benoitc> I don't see the arg about last one
>>>>> [14:27] <papandreou> benoitc: Ah, ok, I understand now. Won't work
very
>>>>> well
>>>>> for me, though, the web app is a single page thing that only asks for
this
>>>>> particular chunk of data once per session, so the ETag will probably
have
>>>>> changed anyway unless we accept day-old data.
>>>>> [14:27] <+benoitc> anyway enotime to discuss about that , i'm on
>>>>> anotherthing
>>>>> [14:32] <papandreou> rnewson: But next step would be for me to
raise the
>>>>> issue on the user mailing list?
>>>>> [14:33] <+rnewson> papandreou: on reflection, it's more a dev@
thing, but
>>>>> yes.
>>>>> [14:33] <+rnewson> post the suggestion about calculating an etag
over the
>>>>> results and then streaming them, with the caveat that a better solution
>>>>> should be found.
>>>>> [14:34] <papandreou> rnewson: Ok, I will, thanks :). Btw. do you
think
>>>>> there's a chance that this will be easier for key=... queries than
>>>>> arbitrary
>>>>> startkey=...&endkey=... ones?
>>>>> [14:35] <+rnewson> papandreou: yes. for key= we could use a bloom
filter.
>>>>> [14:38] <papandreou> rnewson: Man, I've got some reading up to
do :).
>>>>> Thanks! So dev@ it is?
>>>>> [14:39] <+rnewson> papandreou: yes.
>>>>> [14:40] <+rnewson> papandreou: 'bloom filter' is just how we handwave
>>>>> solutions these days, it just sounds vaguely plausible to for the keys=
>>>>> variant
>>>>> [14:40] <+rnewson> but doesn't make sense at all for startkey/endkey
>>>>> [14:40] <+jan____> haha, I'm sitting in an ""HTTP Architecture"
session,
>>>>> and
>>>>> all the two speakers do is tell the audience how CouchDB gets it all
right.
>>>>> [14:41] <+rnewson> at base, we'd want some cheap way to invalidate
a range
>>>>> of keys in memory.
>>>>> [14:49] <+jan____> the answer must include bloom filters.
>>>>>
>>>>
>>>
>>
>

Mime
View raw message