couchdb-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Paul Davis <>
Subject Re: CouchDB 2.0 & out of order changes
Date Sat, 03 Sep 2016 19:52:56 GMT
Hey all,

Bob did a pretty good job explaining how the changes feed works most
of the time but I did want to call attention to an error condition
that might be of concern to the original question. There is a
situation where you can see old changes in the response depending on
some timing and error conditions.

For a bit of background on the since parameter, when we are looking at
a clustered since sequence like such:


What that actually contains is an encoded set of {shard, node,
update_seq} triples kinda like such:

    {"shards/00000000-3fffffff/davisp/test-db.1384769918", db7,  { 9,
<<"ee5754a">>, db7}},
    {"shards/40000000-7fffffff/davisp/test-db.1384769918", db2,  { 1,
<<"0fe9f9c">>, db2}},
    {"shards/80000000-bfffffff/davisp/test-db.1384769918", db5,  {10,
<<"f7b08b9">>, db5}},
    {"shards/c0000000-ffffffff/davisp/test-db.1384769918", db12, {15,
<<"b942877">>, db12}}

What's happening here is that when you specify that parameter, CouchDB
will decode it and then try and read the changes from each of those
shards resuming at the given update seq. As an aside the extra uuid
prefix and node name are extra so that we can try and skip some old
changes, an optimization but not important for this discussion.

Now, the important bit is that if we specify this since seq and one of
the nodes db2, db5, db7, or db12 happens to be down (or just gone if
you stored that since seq for a long time and say the cluster changed
size) then CouchDB has to choose another shard to replace the missing
node. When this happens you will see "old" changes that have already
been processed. Your application should be able to handle this using
the approaches that Bob listed in his email.

However, (there's always a however in distributed systems) there
exists a timing situation where you may be reading the changes feed,
an update comes in to the shard you're reading from, and you see the
change. Then say that node goes down (which would terminate the
changes feed). The client would then reconnect with their latest
update seq and get a different node. This node may (depending on a
whole bunch of timing and optimization things) send a change for the
doc that is technically "before" the change that was already
processed. So you do have to be able to recognize that you already
processed a change for the doc. CouchDB does this internally by
keeping the revision tree and checking revisions against that.

I get the feeling that my description may not be super clear so I'll
try and restate it as a sequence of events:

1. Client requests _changes with since=35-g1AAA...
2. Someone makes an update to doc foo
3. The first shard that handles the update is part of the changes feed from #1
4. Client reading _changes sees update appear in its feed
5. The node containing the shard with an update dies (an operator
rebooted the wrong node perhaps)
6. Client has its changes feed disconnect and reconnects with
since=35-g1AAA.. (or even a newer sequence)
7. The shard that responds as a replacement for the shard on the dead
node does not (yet) have the update
8. The doc exists in the update sequence after where it starts reading
its changes feed
9. Client sees the "old" change and must know that its old

Note that this is all very unlikely and super sensitive to timing. For
instance, one node seeing the update, and then dying, and the other
two copies not receving the update would require some very peculiar
network disconnections in the cluster while still being reachable from
the load balancer. But the possibility remains which means apps will
want to be able to handle it.


On Sat, Sep 3, 2016 at 12:09 PM, Robert Payne <> wrote:
> Thanks for this,
> Sorry it’s a bit light on details, it’s a very specific use case on iOS where cpu/disk
speed is constrained and we have a large data set. The full replication protocol just requires
too many reads/writes to be performant and we’ve optimised it various ways. We’re idempotent
so long as per-document changes are in-order which I was just checking.
> I appreciate the more technical analysis and it certainly clears up what I was asking.
> Cheers,
> Robert
>> On 4/09/2016, at 4:41 AM, Robert Samuel Newson <> wrote:
>> Hi,
>> It is important to understand that the order of rows in the _changes response is
not important. In couchdb before 2.0 the response was totally ordered, but this was never
necessary for correctness. The essential contract for _changes is that you are guaranteed
to see all changes made since the 'since' parameter you pass. The order of those changes is
not guaranteed and it is also not guaranteed that changes from _before_ that 'since' value
are _not_ also returned. The consequence of this contract is that all consumers of the _changes
response must apply each row idempotently. This is true for the replicator, of course.
>> The changes response in 2.0 is partially ordered. The changes from any given shard
will be in a consistent order, but we merge the changes from each shard range of your database
as they are collected from the various contributing nodes, we don't apply a total ordering
over that. The reason is simple; it's expensive and unnecessary. It's important to also remember
that replication, even before 2.0, would not replicate in strict source update order either,
due to (valuable) parallelism when reading changes and applying them.
>> Your question: "Is it possible for the changes feed to send older changes before
newer changes for the same document ID across multiple calls?" requires a little further background
knowledge before answering.
>> While we call it a changes "feed" it's important to remember what it really is, internally,
first. Every database in couchdb, prior to 2.0, is a single file with multiple b+trees recorded
inside it that are kept in absolute sync with each other. One b+tree allows you to look up
a document by the _id parameter. The other b+tree allows you to look up a document by its
update order. It is essential to note that these two b+trees have the same number of key/value
pairs in them at all times.
>> To illustrate this more clearly, consider an empty database. We add one document
to it. It is retrievable by its _id and is also visible in the _changes response as change
number 1. Now, we update that document. It is now change number 2. Change number 1 will never
again appear in the _changes response. That is, every document appears in the _changes response
at its most recent update number.
>> When you call _changes without the continuous parameter, couchdb is simply traversing
that second b+tree and returning each row it finds. It may do this from the beginning (which
was 1 before our update and 2 after) or it may do so from some update seq you supply with
the 'since' parameter.
>> With that now understood, we can look at what changes when we do continuous=true
which is what makes it a "feed" (that is, a potentially unending response of changes as they
are made). This is sent in two phases. The first is exactly as the previous paragraph. Once
all those changes have been sent, couchdb enters a loop where it returns updates as they happen
(or shortly after).
>> It is only in a continuous=true response in couchdb before 2.0 that you would ever
see more than one change for any given document.
>> So, to cut a long story short (too late), the answer to your question is "no". The
changes feed is not a permanent history of all changes made to all documents. Once a document
is updated, it is _moved_ to a newer position and no longer appears in its old one (and no
record of that position is even preserved). Do note, though, that couchdb might return 'Doc
A change (seq: 2-XXXX)' even if your 'since' parameter is _after_ the last change to doc A.
We won't return ' Doc A change (seq: 1-XXXX)' at all after its updated to 2-XXXX.
>> The algorithm for correctly processing the changes response is as follows, and any
variation on this is likely broken;
>> 1) call /_changes?since=0
>> 2) for each returned row, ensure the target has the change in question (either use
_id + _rev to prevent duplicate application of the change or apply the change in a way that
is idempotent)
>> 3) periodically store the update seq of the last processed row to stable storage
(a _local document is a good choice)
>> If you wish to resume applying changes after a shutdown, reboot, or crash, repeat
the above process but substitute your stored update sequence in the ?since= parameter.
>> There are many things that use the changes feed in this way. Within couchdb, there's
database replication (obviously) but also couchdb views. Outside of the core, software like
pouchdb and couchdb-lucene use the changes feed to replicate data or update search indexes.
>> I hope this was useful, and I think it might expose some problems in your couchdb-to-sqlite
synchronisation protocol. Your email is obviously silent on many details there, but if you've
predicated its design on the total ordering properties of couchdb < 2.0, you likely have
some work to do.
>> B.
>>> On 3 Sep 2016, at 00:04, Robert Payne <> wrote:
>>> Hey Everyone,
>>> Reading up on the CouchDB 2.0 migration guides and getting a bit antsy around
the mentions of out of order changes feed and sorts. Is it possible for the changes feed to
send older changes before newer changes for the same document ID across multiple calls?
>>> Assuming start at ?since=“” and always pass in the “last_seq” on every
additional call could a situation like this occur in a single or multiple HTTP calls:
>>> — Changes feed emits Doc A change (seq: 2-XXXX)
>>> — Changes feed emits Doc B change (seq: 3-XXXX)
>>> — Changes feed emits Doc A change (seq: 1-XXXX)
>>> I’m really hoping the case is just that across different doc ids changes can
be out of order. Our use case on mobile is a bit particular as we duplicate edits into a separate
SQLite table and use the changes feed to keep the local database up to date with winning revs
from the server, it just increases the performance of sync by a ton since there is only 1
check and set in SQLite per change that comes in.
>>> Cheers,
>>> Robert

View raw message