On Apr 1, 2009, at 11:03 AM, Chris Anderson wrote:
2) What about _all_docs and seq-num?
I presume _all_docs gets merged like any other view.
_all_docs_by_seq is a
different story. In the current code the sequence number is
incremented by
one for every update. If we want to preserve that behavior in
partitioned
databases we need some sort of consensus algorithm or master
server to
choose the next sequence number. It could easily turn into a
bottleneck or
single point-of-failure if we're not careful.
The alternatives are to a) abandon the current format for update
sequences
in favor of vector clocks or something more opaque, or b) have
_all_docs_by_seq be strictly a node-local query. I'd prefer the
former, as
I think it will be useful for e.g. external indexers to treat the
partitioned database just like a single server one. If we do the
latter, I
think it means that either the external indexers have to be
installed on
every node, or at least they have to be aware of all the partitions.
If at all possible I think we should have the entire partition
group appear
as a single server from the outside. One thing that comes to mind
here is a
question about sequence numbers. Vector clocks only guarantee a
partial
ordering, but I'm under the impression that currently seq numbers
have a
strict ordering.
Database sequence numbers are used in replication and in
determining whether
views need refreshing. Anything else I'm missing? Currently it
seems there
is no tracking of which updates actually change a view index
(comment on
line 588 of couch_httpd_view.erl on trunk). Improving this would be
a nice
win. See my answer to number (3).
The easy way to manage seq numbers is to let one node be the write
master
for any cluster. (The root node of any partition group could
actually be a
cluster, but if writes always go through a master the master can
maintain
the global sequence number for the partition group).
The problem with this approach is that the main use-case for
partitioning is when your incoming writes exceed the capacity of a
single node. By partitioning the key-space, you can get more
write-throughput.
I think Randall was saying requests just have to originate at the
master node. That master node could do nothing more than assign a
sequence number, choose a node, and proxy the request down the tree
for the heavy lifting. I bet we could get pretty good throughput, but
I still worry about this approach for availability reasons.
I'm not sure that an update-seq per node is such a bad thing, as it
will require any external indexers to be deployed in a 1-to-1
relationship to nodes, which automatically balances the load for the
indexer as well. With a merged seq-id, users would be encouraged to
partition CouchDB without bothering to partition indexers. Maybe this
is acceptable in some cases, but not in the general case.
So, the vector clock approach still has a per-node update sequence for
each node's local clock, it just does the best job possible of
globally ordering those per-node sequences. We could easily offer
local update sequences as well via some query string parameter. I
understand the desire to encourage partitioned indexers, but I believe
that won't always be possible. Bottom line, I think we should support
global indexing of a partitioned DB.
One other thing that bothers me is the merge-sort required for
every view
lookup. In *really* large clusters it won't be good if queries
for a single
key in a view have to hit each partition. We could have an
alternative
structure where each view gets partitioned much like the document
data while
its built. I worry that a view partitioned in this way may need
frequent
rebalancing during the build, since view keys are probably not
going to be
uniformly distributed. Nevertheless, I think the benefit of
having many
view queries only hit a small subset of nodes in the cluster is
pretty huge.
I agree that the merge-sort is something we need to look at
carefully. We
should never hit a node in a view query unless it has data we need.
We
certainly can't avoid merging altogether, but we can make an effort
to do
smart rebalancing later on.
I think rebalancing aka shuffling will turn out to be one of those
devil-in-the-details things. Because any document can emit any key, in
the case of rebalancing, if you have to rebuild part of an index due
to node-failure, you'd need to re-request from every other node, any
view rows that might fit in that range. This requires every node to
know about every other node.
If view data is stored with document data, then nodes need only know
about their child nodes, in the tree structure. Recovering from
node-failure is easy: just swap in the failed node's hot-backup, and
regenerate the views on it.
I agree that the cost of merge sort will be ongoing, but I think the
simplicity of this approach at least indicates that we should take it
for the initial work. If we consider rebalancing an optimization, we
can add it later.
I think a better optimization would be to have inner nodes of the tree
lazily cache the view rows of their children. This way the computation
is spread out but the hops for popular queries can be mostly
eliminated.
Ok, +1 from me. I totally agree that rebalancing can get hairy, but
hey, that's what makes this fun!
4) Should the consistent hashing algorithm map ids to leaf nodes
or just
to
children? I lean toward children because it encapsulates
knowledge about
the
layout of subtrees at each tree level.
If the algorithm maps to children, does that mean every document
lookup has
to traverse the tree? I'm not sure that's a great idea. Up to
~100 nodes I
think it may be better to have all document lookups take O(1)
hops. I think
distributed Erlang can keep a system of that size globally
connected without
too much trouble.
I like the strict tree approach. I'd translate Adam's comment as:
distributed Erlang can probably handle a tree of depth=1, even with
~100 nodes.
I'd like to hear more about how we implement redundancy and handle
node failures in the tree structure. In a pure consistent hashing
ring, whether globally connected (Dynamo) or not (Chord), there are
clear procedures for dealing with node failures, usually involving
storing copies of the data at adjacent nodes along the ring. Do we
have an analogue of that in the tree? I'm especially worried about
what happens when inner nodes go down.
Best, Adam
