give proxied cluster nodes a higher cost than the cluster gateway

This makes eg git-annex get default to using the cluster rather than an
arbitrary node, which is better UI.

The actual cost of accessing a proxied node vs using the cluster is
basically the same. But using the cluster allows smarter load-balancing
to be done on the cluster.

diff --git a/Remote/Git.hs b/Remote/Git.hs
index 89e0da38c1bc3179d56ceec238318d353a219637..844beb3b2e70c04f73868175d08925238a7bc06e 100644 (file)
--- a/Remote/Git.hs
+++ b/Remote/Git.hs
@@ -176,6 +176,7 @@ configRead autoinit r = do
                        Just r' -> return r'
                _ -> return r
 
+
 gen :: Git.Repo -> UUID -> RemoteConfig -> RemoteGitConfig -> RemoteStateHandle -> Annex (Maybe Remote)
 gen r u rc gc rs
        -- Remote.GitLFS may be used with a repo that is also encrypted
@@ -186,10 +187,9 @@ gen r u rc gc rs
                Nothing -> do
                        st <- mkState r u gc
                        c <- parsedRemoteConfig remote rc
-                       go st c <$> remoteCost gc c defcst
+                       go st c <$> remoteCost gc c (defaultRepoCost r)
                Just addr -> Remote.P2P.chainGen addr r u rc gc rs
   where
-       defcst = if repoCheap r then cheapRemoteCost else expensiveRemoteCost
        go st c cst = Just new
          where
                new = Remote 
@@ -229,6 +229,11 @@ gen r u rc gc rs
                        , remoteStateHandle = rs
                        }
 
+defaultRepoCost :: Git.Repo -> Cost
+defaultRepoCost r
+       | repoCheap r = cheapRemoteCost
+       | otherwise = expensiveRemoteCost
+
 unavailable :: Git.Repo -> UUID -> RemoteConfig -> RemoteGitConfig -> RemoteStateHandle -> Annex (Maybe Remote)
 unavailable r = gen r'
   where
@@ -854,12 +859,17 @@ listProxied proxies rs = concat <$> mapM go rs
                -- that cluster does not need to be synced with
                -- by default, because syncing with the cluster will
                -- effectively sync with all of its nodes.
+               --
+               -- Also, give it a slightly higher cost than the
+               -- cluster by default, to encourage using the cluster.
                adjustclusternode clusters =
                        case M.lookup (ClusterNodeUUID (proxyRemoteUUID p)) (clusterNodeUUIDs clusters) of
                                Just cs
                                        | any (\c -> S.member (fromClusterUUID c) proxieduuids) (S.toList cs) ->
                                                addremoteannexfield SyncField
                                                        [Git.ConfigValue $ Git.Config.boolConfig' False]
+                                               . addremoteannexfield CostField 
+                                                       [Git.ConfigValue $ encodeBS $ show $ defaultRepoCost r + 0.1]
                                _ -> id
 
                proxieduuids = S.map proxyRemoteUUID proxied

diff --git a/doc/todo/git-annex_proxies.mdwn b/doc/todo/git-annex_proxies.mdwn
index a53197f8c1fa0dad0efe02dec872162695363ff3..29f662e6bf86ba138493cd40ffbbdb7a4eac76b5 100644 (file)
--- a/doc/todo/git-annex_proxies.mdwn
+++ b/doc/todo/git-annex_proxies.mdwn
@@ -41,14 +41,17 @@ For June's work on [[design/passthrough_proxy]], remaining todos:
   eg prefer to avoid using remotes that are doing other transfers at the
   same time.
 
-* The cost of a cluster and of its proxied nodes is currently all the same.
-  It would make sense for proxied nodes that are accessed via an intermedia
-  gateway to have a higher cost than proxied nodes that are accessed via
-  the remote gateway. And proxied nodes should generally have a higher cost
-  than the cluster, so that git-annex defaults to using the cluster.
-  (The cost of accessing a proxied node vs using the cluster is the same,
-  but using the cluster allows smarter load-balancing to be done on the
-  cluster. It also makes the UI not mention individual nodes.)
+* The cost of a proxied node that is accessed via an intermediate gateway
+  is currently the same as a node accessed via the cluster gateway.
+  To fix this, there needs to be some way to tell how many hops through
+  gateways it takes to get to a node. Currently the only way is to
+  guess based on number of dashes in the node name, which is not satisfying.
+
+  Even counting hops is not very satisfying, one cluster gateway could
+  be much more expensive to traverse than another one.
+
+  If seriously tackling this, it might be worth making enough information
+  available to use spanning tree protocol for routing inside clusters.
 
 * Optimise proxy speed. See design for ideas.
 
@@ -117,3 +120,6 @@ For June's work on [[design/passthrough_proxy]], remaining todos:
   protocol messages on to any remotes that have the same UUID as
   the cluster. Needs extension to P2P protocol to avoid cycles.
   (done)
+
+* Proxied cluster nodes should have slightly higher cost than the cluster
+  gateway. (done)