incubator-couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From J Chris Anderson <>
Subject Re: Large lists of data
Date Sat, 24 Jul 2010 19:49:49 GMT

On Jul 24, 2010, at 12:35 PM, John wrote:

> Hi
> Thanks to you both for the answers so far. Indeed my setup is far more complex than I
have exposed to date but I'm making it into bite sized chunks around the Use Cases that I
think are the more challenging for me.
> Although your answers were useful they don't quite hit the mark and that's probably because
I didn't explain my problem well enough to start with!
> The database will contain entries from multiple lists (many thousands perhaps) so the
_id will never be unique on a telephone number. Perhaps this might work though:
> GET /database/<list _id>#0123456789
> or I could just keep the _id as a uuid and move this problem (find by list id and number)
to the view.

for now I'd say just go with uuids and you can have a view by telephone number for direct
(or starts_with) lookups.

> The view by list wont work for me. I need to be able to query the view with something
> GET  /database/_design/portability/_view/NP?key=0123456789&list=<_id of list>
> In fact in some cases the problem is more complex than this as I need to search for "widest
> GET  /database/_design/portability/_view/NP?key=0123456789&list=<_id of list>&min_width=5
> which would return the widest match in:
> 0123456789
> 012345678
> 01234567
> 0123456
> 012345
> 01234
> I even have another use case where I need to do a STARTS_WITH e.g. provide a key of 01234
and return true if there are any numbers that start 01234.

this is easy. have a view like:

function(doc) {
  emit([doc.list_id, doc.number], null) 

Then you can query with 

?startkey=["mylist", "012"]&endkey=["mylist", "013"]

to get everything with a prefix of "012" in the "mylist" list. you can mess around with the
endkey_inclusive (or is it inclusive_endkey) = true / false to not get the exact number "013"
in your result set.

from this technique you can see how you could do starts-with against just phone numbers also,
with a view like

function(doc_ {
  emit(doc.number, null)

Note I have telephone numbers as strings in this example as a regular number 012 is the same
as 12.

> This is a typical telecom problem and it would be good to document a Design Pattern for
this Use Case. In fact there's a discussion for another day on how/where we could document
this patterns and get peer reviews on them.
> Thanks again
> John
> On 24 Jul 2010, at 19:15, J Chris Anderson wrote:
>> On Jul 24, 2010, at 7:41 AM, wrote:
>>> Hello,
>>> 1/ it's a little hard to answer this question, your setup is certainly a little
more complex than what you expose in your email :-) However thousands of documents are gracefuly
handled by CouchDB.
>>> 2/ At first sight your documents will look like :
>>> { "_id": 0123456789 , "list": "mylist", "type": "NP", "status":"portedIn", "operatorId":1234
>>> That way you can query your document by phone number :
>>> GET /database/0123456789
>>> and have all documents belonging to the list "mylist" by creating a view that
emits the "list" field :
>>> function (doc) {
>>> if ( doc.list  && doc.type == "NP" ) {
>>>  emit (doc.list,null);
>>> }
>>> }
>>> and fetching them with something like :
>>> GET /database/_design/portability/_view/NP?key="mylist"&include_docs=true
>>> 3/ When updating a document : the document is of course immediately available.
However the view index won't be updated. In CouchDB view indexes are rebuilt on view query
(not on document update). When you'll query CouchDB "give me all the documents of the view
NP", Couch will take all documents that have changed (added, updated, deleted) since the last
time you asked Couch for the view, and will update indexes accordingly. You have the option
of fetching the view without rebuilding the index, with the "stale" parameter, but in this
case, of course, you won't see the changes. During the rebuilt of the index, subsequent view
queries are queued until the index is up to date.
>>> 4/ I setup CouchDB to parse network logs. A view took something like 25 minuts
for 100 millions documents, on a Dell PowerEdge 2950 Xen Virtual Machine with two dedicated
processors and 4gigs ram. Numbers can heavily vary according to the complexity of the view,
so it's always hard (and dangerous) to give numbers. Moreover my indexes were not only numbers,
but also strings.
>> this is a good response. I'd only follow up to say that there are some techniques
you can use to further tune view-generation performance. one: keysize and entropy can make
a big difference. the view by list, as above, looks pretty good on that front.
>> CouchDB can also be configured to store view indexes on a separate disk from the
database file, which can reduce IO contention if you are at the edge of what your hardware
can do.
>> Also, there is the option to query views with stale=ok, which will return a query
based on the latest snapshot, with low latency, so clients aren't blocked waiting for generation
to complete. then you can use a cron-job with a regular view query and limit=1 to keep the
index up to date. so clients always see a fairly recent snapshot, with low latency.
>>> What you should be aware of is that CouchDB requires maintenance tasks to keep
great performances, it's called "compact" and should be run on databases (to rebuilt the db
file that is append-only) and on databases views (to rebuild the index file that is append-only).
During the compact, database is still available but performances are degraded (from my personnal
>>> Also, a new replication engine is in the pipe and should greatly improve the
replication experience.
>>> Mickael
>>> ----- Mail Original -----
>>> De: "John" <>
>>> À:
>>> Envoyé: Samedi 24 Juillet 2010 11h37:56 GMT +01:00 Amsterdam / Berlin / Berne
/ Rome / Stockholm / Vienne
>>> Objet: Large lists of data
>>> Hi 
>>> I'm currently evaluating couchdb as a candidate to replace the relational databases
as used in our Telecom Applications.
>>> For most of our data I can see a good fit and we already expose our service provisioning
as json over REST so we're well positioned for a migration.
>>> One area that concerns me though is whether this technology is suitable for our
list data. An example of this is Mobile Number Portability where we have millions of rows
of data representing ported numbers with some atrributes against each.
>>> We use the standard Relational approach to this and have an entries table that
has a foreign key reference to a parent list. 
>>> On our web services we do something like this:
>>> Create a List:
>>> PUT /cie-rest/provision/accounts/netdev/lists/mylist
>>> { "type": "NP"}
>>> To add a row to a list 
>>> PUT /cie-rest/provision/accounts/netdev/lists/mylist/entries/0123456789
>>> { "status":"portedIn", "operatorId":1234}
>>> If we want to add a lot of rows we just POST a document to the list.
>>> The list data is used when processing calls and it requires a fast lookup on
the entries table which is obviously indexed.
>>> Anyway, I'd be interested in getting some opinions on:
>>> 1) Is couchdb the *right* technology for this job? (I know it can do it!)
>>> 2) I presume that the relationship I currently have in my relational database
would remain the same for couch i.e. The entry document would ref the list document but maybe
there's a better way to do this?
>>> 3) Number portability requires 15 min, 1 hour and daily syncs with a central
number portability database. This can result in bulk updates of thousands of numbers. I'm
concerned with how long it takes to build a couchdb index and to incrementally update it when
the number of changes is large (Adds/removes).  
>>> What does this mean to the availability of the number? i.e. Is the entry in the
db but its unavailable to the application as it's entry in the index hasnt been built yet?
>>> 4) Telephone numbers like btrees so the index building should be quite fast and
efficient I would of thought but does someone have anything more concrete in terms of how
long it would take typically? I think that the bottleneck is the disk i/o and therefore it
may be vastly different between my laptop and one of our beefy production servers but again
I'd be interested in other peoples experience.
>>> Bit of a long one so thanks if you've read it to this point! There's a lot to
like with couchdb (esp the replication for our use case) so I'm hoping that what i've asked
above is feasible!
>>> Thanks
>>> John

View raw message