couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Matt Goodall <matt.good...@gmail.com>
Subject Re: Deleted docs can be accidentally resurrected
Date Fri, 20 Nov 2009 10:01:00 GMT
2009/11/20 Matt Goodall <matt.goodall@gmail.com>:
> 2009/11/19 Adam Kocoloski <kocolosk@apache.org>:
>> On Nov 18, 2009, at 6:51 PM, Matt Goodall wrote:
>>
>>> 2009/11/18 Sebastian Cohnen <sebastiancohnen@googlemail.com>:
>>>> afaik, this behavior is normal. you just created another (and new) resource/document.
>>>>
>>>> have a look (the database test is new and never contained docs with id 'foo'
or 'foobar'):
>>>>
>>>> $ curl -X PUT -d '{"_id": "foo", "_rev": "1-967a00dff5e02add41819138abb3284d",
"num": 1}' "http://localhost:5984/test/foo"
>>>> {"ok":true,"id":"foo","rev":"2-65371ad05fc99a9b794f68c6003bc8da"}
>>>>
>>>> $ curl -X PUT -d '{"_id": "foobar", "_rev": "4-anythingyouwant", "num": 1}'
"http://localhost:5984/test/foobar"
>>>> {"ok":true,"id":"foobar","rev":"5-084106c9189a346875996e76c1833630"}
>>>
>>> Hmm, seems wrong to me although I can imagine that's useful during
>>> replication (which doesn't mean it should work for a PUT). However,
>>> why does the following fail then:
>>>
>>> $ curl -X "PUT" -d '{}' "http://localhost:15984/test/foo"
>>> {"ok":true,"id":"foo","rev":"1-967a00dff5e02add41819138abb3284d"}
>>> $ curl -X "PUT" -d '{"_id": "foo", "_rev":
>>> "1-967a00dff5e02add41819138abb3284d"}'
>>> "http://localhost:15984/test/foo"
>>> {"ok":true,"id":"foo","rev":"2-7051cbe5c8faecd085a3fa619e6e6337"}
>>> $ curl -X "DELETE"
>>> "http://localhost:15984/test/foo?rev=2-7051cbe5c8faecd085a3fa619e6e6337"
>>> {"ok":true,"id":"foo","rev":"3-7379b9e515b161226c6559d90c4dc49f"}
>>> $ curl -X "PUT" -d '{"_id": "foo", "_rev":
>>> "1-967a00dff5e02add41819138abb3284d"}'
>>> "http://localhost:15984/test/foo"
>>> {"error":"conflict","reason":"Document update conflict."}
>>>
>>> - Matt
>>>
>>>>
>>>>
>>>> cheers,
>>>>
>>>> tisba / Sebastian
>>>>
>>>> On 18.11.2009, at 23:56, Matt Goodall wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> I swear I've reported this odd behaviour before but I can't find any
>>>>> mention of it now ...
>>>>>
>>>>> I have some code with a race condition that is caused by what I
>>>>> believe is a bug in CouchDB - a deleted document can be updated using
>>>>> an old, valid rev.
>>>>>
>>>>> If a document has not been deleted then CouchDB correctly returns a
>>>>> conflict error if the latest rev is not sent. However, once deleted,
>>>>> any rev in the docs history can be sent as the update and, as long as
>>>>> the doc is changed in some way, the document will be resurrected with
>>>>> a rev whose sequence is 1 more than that sent in the update.
>>>>>
>>>>> Create new doc ...
>>>>> $ curl -X "PUT" -d '{"_id": "foo"}' http://localhost:15984/test/foo
>>>>> {"ok":true,"id":"foo","rev":"1-967a00dff5e02add41819138abb3284d"}
>>>>>
>>>>> Update a couple of times to move the rev on ...
>>>>> $ curl -X "PUT" -d '{"_id": "foo", "_rev":
>>>>> "1-967a00dff5e02add41819138abb3284d"}' http://localhost:15984/test/foo
>>>>> {"ok":true,"id":"foo","rev":"2-7051cbe5c8faecd085a3fa619e6e6337"}
>>>>> $ curl -X "PUT" -d '{"_id": "foo", "_rev":
>>>>> "2-7051cbe5c8faecd085a3fa619e6e6337"}' http://localhost:15984/test/foo
>>>>> {"ok":true,"id":"foo","rev":"3-825cb35de44c433bfb2df415563a19de"}
>>>>>
>>>>> Delete the doc ...
>>>>> $ curl -X "DELETE"
>>>>> "http://localhost:15984/test/foo?rev=3-825cb35de44c433bfb2df415563a19de"{"ok":true,"id":"foo","rev":"4-1df13287548620bf858cf9d1b810972a"}
>>>>>
>>>>> Update using an old rev, changing something ...
>>>>> $ curl -X "PUT" -d '{"_id": "foo", "_rev":
>>>>> "1-967a00dff5e02add41819138abb3284d", "num": 1}'
>>>>> http://localhost:15984/test/foo
>>>>> {"ok":true,"id":"foo","rev":"2-65371ad05fc99a9b794f68c6003bc8da"}
>>>>>
>>>>> Now, I can sort of understand how something like that might happen
>>>>> during replication but there's no replication going on here and
>>>>> there's no conflict created in the database.
>>>>>
>>>>> I can't think of a way around the problem at the moment, other than
>>>>> marking the document for deletion and then sweeping it later when
>>>>> something else is hopefully not going to be touching the document.
>>>>> But, frankly, that's horrible,
>>>>>
>>>>> - Matt
>>
>> Hi Matt, bizarre.  Your latest email shows the behavior I expect -- you tried to
PUT a document using an incorrect MVCC revision, and you got a conflict as a result.  A successful
PUT of a new document with an _id that corresponds to a previously deleted document should
omit the MVCC rev.
>
> Yep, I only posted that (it's what I would expect too) to demonstrate
> how weird the problem is.
>
>>
>> Your first email is the weird one -- you supplied an old MVCC _rev, but changed the
body to something different than the body at the time the doc was deleted, and it worked.
 I can confirm this.  Can you file a JIRA ticket?  Best,
>
> Exactly. If there's a rev then it should match the current rev of the
> doc or fail.
>
> I'll file a ticket tomorrow morning.

There's a ticket for it already actually,
https://issues.apache.org/jira/browse/COUCHDB-292. The ticket also
contains a still failing test.

- Matt

Mime
View raw message