riak_test/tests/replication.erl
2014-04-07 18:06:57 +00:00

704 lines
25 KiB
Erlang

-module(replication).
-behavior(riak_test).
-export([confirm/0]).
-compile(export_all).
-include_lib("eunit/include/eunit.hrl").
-import(rt, [deploy_nodes/2,
join/2,
wait_until_nodes_ready/1,
wait_until_no_pending_changes/1]).
%% export functions shared with other replication tests...
-export([make_bucket/3]).
confirm() ->
Conf = [
{riak_repl,
[
{fullsync_on_connect, false},
{fullsync_interval, disabled},
{diff_batch_size, 10}
]}
],
rt:set_advanced_conf(all, Conf),
[ANodes, BNodes] = rt:build_clusters([3, 3]),
replication(ANodes, BNodes, false),
pass.
replication([AFirst|_] = ANodes, [BFirst|_] = BNodes, Connected) ->
AllNodes = ANodes ++ BNodes,
rt:log_to_nodes(AllNodes, "Starting replication test"),
TestHash = erlang:md5(term_to_binary(os:timestamp())),
TestBucket = <<TestHash/binary, "-systest_a">>,
FullsyncOnly = <<TestHash/binary, "-fullsync_only">>,
RealtimeOnly = <<TestHash/binary, "-realtime_only">>,
NoRepl = <<TestHash/binary, "-no_repl">>,
case Connected of
false ->
%% clusters are not connected, connect them
%% write some initial data to A
lager:info("Writing 100 keys to ~p", [AFirst]),
?assertEqual([], do_write(AFirst, 1, 100, TestBucket, 2)),
rt:log_to_nodes(AllNodes, "Adding listeners"),
%% setup servers/listeners on A
Listeners = add_listeners(ANodes),
rt:wait_until_ring_converged(ANodes),
%% verify servers are visible on all nodes
verify_listeners(Listeners),
lager:info("waiting for leader to converge on cluster A"),
?assertEqual(ok, wait_until_leader_converge(ANodes)),
lager:info("waiting for leader to converge on cluster B"),
?assertEqual(ok, wait_until_leader_converge(BNodes)),
%% get the leader for the first cluster
LeaderA = rpc:call(AFirst, riak_repl_leader, leader_node, []),
%% list of listeners not on the leader node
NonLeaderListeners = lists:keydelete(LeaderA, 3, Listeners),
rt:log_to_nodes(AllNodes, "Setup replication sites"),
%% setup sites on B
%% TODO: make `NumSites' an argument
NumSites = 4,
{Ip, Port, _} = hd(NonLeaderListeners),
add_site(hd(BNodes), {Ip, Port, "site1"}),
rt:wait_until_ring_converged(BNodes),
FakeListeners = gen_fake_listeners(NumSites-1),
add_fake_sites(BNodes, FakeListeners),
rt:wait_until_ring_converged(BNodes),
rt:log_to_nodes(AllNodes, "Verify replication sites"),
%% verify sites are distributed on B
verify_sites_balanced(NumSites, BNodes),
wait_until_connection(LeaderA),
%% check the listener IPs were all imported into the site
verify_site_ips(BFirst, "site1", Listeners);
_ ->
lager:info("waiting for leader to converge on cluster A"),
?assertEqual(ok, wait_until_leader_converge(ANodes)),
lager:info("waiting for leader to converge on cluster B"),
?assertEqual(ok, wait_until_leader_converge(BNodes)),
%% get the leader for the first cluster
LeaderA = rpc:call(AFirst, riak_repl_leader, leader_node, []),
lager:info("Leader on cluster A is ~p", [LeaderA]),
[{Ip, Port, _}|_] = get_listeners(LeaderA)
end,
rt:log_to_nodes(AllNodes, "Write data to A"),
%% write some data on A
?assertEqual(ok, wait_until_connection(LeaderA)),
%io:format("~p~n", [rpc:call(LeaderA, riak_repl_console, status, [quiet])]),
lager:info("Writing 100 more keys to ~p", [LeaderA]),
?assertEqual([], do_write(LeaderA, 101, 200, TestBucket, 2)),
rt:log_to_nodes(AllNodes, "Verify data received on B"),
%% verify data is replicated to B
lager:info("Reading 100 keys written to ~p from ~p", [LeaderA, BFirst]),
?assertEqual(0, wait_for_reads(BFirst, 101, 200, TestBucket, 2)),
case Connected of
false ->
%% check that the keys we wrote initially aren't replicated yet, because
%% we've disabled fullsync_on_connect
lager:info("Check keys written before repl was connected are not present"),
Res2 = rt:systest_read(BFirst, 1, 100, TestBucket, 2),
?assertEqual(100, length(Res2)),
start_and_wait_until_fullsync_complete(LeaderA),
lager:info("Check keys written before repl was connected are present"),
?assertEqual(0, wait_for_reads(BFirst, 1, 200, TestBucket, 2));
_ ->
ok
end,
ASecond = hd(ANodes -- [LeaderA]),
%% disconnect the other cluster, so realtime doesn't happen
lager:info("disconnect the 2 clusters"),
del_site(BNodes, "site1"),
?assertEqual(ok, wait_until_no_connection(LeaderA)),
lager:info("write 2000 keys"),
?assertEqual([], do_write(ASecond, 50000, 52000,
TestBucket, 2)),
lager:info("reconnect the 2 clusters"),
add_site(hd(BNodes), {Ip, Port, "site1"}),
?assertEqual(ok, wait_until_connection(LeaderA)),
start_and_wait_until_fullsync_complete(LeaderA),
lager:info("read 2000 keys"),
?assertEqual(0, wait_for_reads(BFirst, 50000, 52000, TestBucket, 2)),
%%
%% Failover tests
%%
rt:log_to_nodes(AllNodes, "Testing master failover: stopping ~p", [LeaderA]),
lager:info("Testing master failover: stopping ~p", [LeaderA]),
rt:stop(LeaderA),
rt:wait_until_unpingable(LeaderA),
wait_until_leader(ASecond),
LeaderA2 = rpc:call(ASecond, riak_repl_leader, leader_node, []),
lager:info("New leader is ~p", [LeaderA2]),
?assertEqual(ok, wait_until_connection(LeaderA2)),
lager:info("Writing 100 more keys to ~p now that the old leader is down",
[ASecond]),
?assertEqual([], do_write(ASecond, 201, 300, TestBucket, 2)),
%% verify data is replicated to B
lager:info("Reading 100 keys written to ~p from ~p", [ASecond, BFirst]),
?assertEqual(0, wait_for_reads(BFirst, 201, 300, TestBucket, 2)),
%% get the leader for the first cluster
LeaderB = rpc:call(BFirst, riak_repl_leader, leader_node, []),
lager:info("Testing client failover: stopping ~p", [LeaderB]),
rt:stop(LeaderB),
rt:wait_until_unpingable(LeaderB),
BSecond = hd(BNodes -- [LeaderB]),
wait_until_leader(BSecond),
LeaderB2 = rpc:call(BSecond, riak_repl_leader, leader_node, []),
lager:info("New leader is ~p", [LeaderB2]),
?assertEqual(ok, wait_until_connection(LeaderA2)),
lager:info("Writing 100 more keys to ~p now that the old leader is down",
[ASecond]),
?assertEqual([], do_write(ASecond, 301, 400, TestBucket, 2)),
%% verify data is replicated to B
rt:wait_until_pingable(BSecond),
lager:info("Reading 101 keys written to ~p from ~p", [ASecond, BSecond]),
?assertEqual(0, wait_for_reads(BSecond, 301, 400, TestBucket, 2)),
%% Testing fullsync with downed nodes
lager:info("Re-running fullsync with ~p and ~p down", [LeaderA, LeaderB]),
start_and_wait_until_fullsync_complete(LeaderA2),
%%
%% Per-bucket repl settings tests
%%
lager:info("Restarting down node ~p", [LeaderA]),
rt:start(LeaderA),
rt:wait_until_pingable(LeaderA),
wait_until_no_pending_changes(ANodes),
wait_until_leader_converge(ANodes),
start_and_wait_until_fullsync_complete(LeaderA2),
case nodes_all_have_version(ANodes, "1.2.2") of
true ->
lager:info("Starting Joe's Repl Test"),
%% At this point, realtime sync should still work, but, it doesn't
%% because of a bug in 1.2.1.
%% Check that repl leader is LeaderA
%% Check that LeaderA2 has ceeded socket back to LeaderA
lager:info("Leader: ~p", [rpc:call(ASecond, riak_repl_leader, leader_node, [])]),
lager:info("LeaderA: ~p", [LeaderA]),
lager:info("LeaderA2: ~p", [LeaderA2]),
?assertEqual(ok, wait_until_connection(LeaderA)),
lager:info("Simulation partition to force leader re-election"),
OldCookie = rpc:call(LeaderA2, erlang, get_cookie, []),
NewCookie = list_to_atom(lists:reverse(atom_to_list(OldCookie))),
rpc:call(LeaderA2, erlang, set_cookie, [LeaderA2, NewCookie]),
[ rpc:call(LeaderA2, erlang, disconnect_node, [Node]) ||
Node <- ANodes -- [LeaderA2]],
[ rpc:call(Node, erlang, disconnect_node, [LeaderA2]) ||
Node <- ANodes -- [LeaderA2]],
wait_until_new_leader(hd(ANodes -- [LeaderA2]), LeaderA2),
InterimLeader = rpc:call(LeaderA, riak_repl_leader, leader_node, []),
lager:info("Interim leader: ~p", [InterimLeader]),
rpc:call(LeaderA2, erlang, set_cookie, [LeaderA2, OldCookie]),
[ rpc:call(LeaderA2, net_adm, ping, [Node]) ||
Node <- ANodes -- [LeaderA2]],
[ rpc:call(Node, net_adm, ping, [LeaderA2]) ||
Node <- ANodes -- [LeaderA2]],
%% there's no point in writing anything until the leaders
%% converge, as we can drop writes in the middle of an election
wait_until_leader_converge(ANodes),
LeaderA3 = rpc:call(ASecond, riak_repl_leader, leader_node, []),
wait_until_connection(LeaderA3),
lager:info("Leader: ~p", [LeaderA3]),
lager:info("Writing 2 more keys to ~p", [LeaderA3]),
?assertEqual([], do_write(LeaderA3, 1301, 1302, TestBucket, 2)),
%% verify data is replicated to B
lager:info("Reading 2 keys written to ~p from ~p", [LeaderA3, BSecond]),
?assertEqual(0, wait_for_reads(BSecond, 1301, 1302, TestBucket, 2)),
lager:info("Finished Joe's Section"),
lager:info("Nodes restarted");
_ ->
lager:info("Skipping Joe's Repl Test")
end,
lager:info("Restarting down node ~p", [LeaderB]),
rt:start(LeaderB),
rt:wait_until_pingable(LeaderB),
case nodes_all_have_version(ANodes, "1.1.0") of
true ->
make_bucket(ANodes, NoRepl, [{repl, false}]),
case nodes_all_have_version(ANodes, "1.2.0") of
true ->
make_bucket(ANodes, RealtimeOnly, [{repl, realtime}]),
make_bucket(ANodes, FullsyncOnly, [{repl, fullsync}]),
%% disconnect the other cluster, so realtime doesn't happen
lager:info("disconnect the 2 clusters"),
del_site(BNodes, "site1"),
?assertEqual(ok, wait_until_no_connection(LeaderA)),
lager:info("write 100 keys to a realtime only bucket"),
?assertEqual([], do_write(ASecond, 1, 100,
RealtimeOnly, 2)),
lager:info("reconnect the 2 clusters"),
add_site(LeaderB, {Ip, Port, "site1"}),
?assertEqual(ok, wait_until_connection(LeaderA));
_ ->
timer:sleep(1000)
end,
LeaderA4 = rpc:call(ASecond, riak_repl_leader, leader_node, []),
lager:info("write 100 keys to a {repl, false} bucket"),
?assertEqual([], do_write(ASecond, 1, 100, NoRepl, 2)),
case nodes_all_have_version(ANodes, "1.2.0") of
true ->
lager:info("write 100 keys to a fullsync only bucket"),
?assertEqual([], do_write(ASecond, 1, 100,
FullsyncOnly, 2)),
lager:info("Check the fullsync only bucket didn't replicate the writes"),
Res6 = rt:systest_read(BSecond, 1, 100, FullsyncOnly, 2),
?assertEqual(100, length(Res6)),
lager:info("Check the realtime only bucket that was written to offline "
"isn't replicated"),
Res7 = rt:systest_read(BSecond, 1, 100, RealtimeOnly, 2),
?assertEqual(100, length(Res7));
_ ->
timer:sleep(1000)
end,
lager:info("Check the {repl, false} bucket didn't replicate"),
Res8 = rt:systest_read(BSecond, 1, 100, NoRepl, 2),
?assertEqual(100, length(Res8)),
%% do a fullsync, make sure that fullsync_only is replicated, but
%% realtime_only and no_repl aren't
start_and_wait_until_fullsync_complete(LeaderA4),
case nodes_all_have_version(ANodes, "1.2.0") of
true ->
lager:info("Check fullsync only bucket is now replicated"),
?assertEqual(0, wait_for_reads(BSecond, 1, 100,
FullsyncOnly, 2)),
lager:info("Check realtime only bucket didn't replicate"),
Res10 = rt:systest_read(BSecond, 1, 100, RealtimeOnly, 2),
?assertEqual(100, length(Res10)),
lager:info("Write 100 more keys into realtime only bucket"),
?assertEqual([], do_write(ASecond, 101, 200,
RealtimeOnly, 2)),
timer:sleep(5000),
lager:info("Check the realtime keys replicated"),
?assertEqual(0, wait_for_reads(BSecond, 101, 200,
RealtimeOnly, 2)),
lager:info("Check the older keys in the realtime bucket did not replicate"),
Res12 = rt:systest_read(BSecond, 1, 100, RealtimeOnly, 2),
?assertEqual(100, length(Res12));
_ ->
ok
end,
lager:info("Check {repl, false} bucket didn't replicate"),
Res13 = rt:systest_read(BSecond, 1, 100, NoRepl, 2),
?assertEqual(100, length(Res13));
_ ->
ok
end,
lager:info("Test passed"),
fin.
verify_sites_balanced(NumSites, BNodes0) ->
Leader = rpc:call(hd(BNodes0), riak_repl_leader, leader_node, []),
case node_has_version(Leader, "1.2.0") of
true ->
BNodes = nodes_with_version(BNodes0, "1.2.0") -- [Leader],
NumNodes = length(BNodes),
case NumNodes of
0 ->
%% only leader is upgraded, runs clients locally
?assertEqual(NumSites, client_count(Leader));
_ ->
NodeCounts = [{Node, client_count(Node)} || Node <- BNodes],
lager:notice("nodecounts ~p", [NodeCounts]),
lager:notice("leader ~p", [Leader]),
Min = NumSites div NumNodes,
[?assert(Count >= Min) || {_Node, Count} <- NodeCounts]
end;
false ->
ok
end.
%% does the node meet the version requirement?
node_has_version(Node, Version) ->
{_, NodeVersion} = rpc:call(Node, init, script_id, []),
case NodeVersion of
current ->
%% current always satisfies any version check
true;
_ ->
NodeVersion >= Version
end.
nodes_with_version(Nodes, Version) ->
[Node || Node <- Nodes, node_has_version(Node, Version)].
nodes_all_have_version(Nodes, Version) ->
Nodes == nodes_with_version(Nodes, Version).
client_count(Node) ->
Clients = rpc:call(Node, supervisor, which_children, [riak_repl_client_sup]),
length(Clients).
gen_fake_listeners(Num) ->
Ports = gen_ports(11000, Num),
IPs = lists:duplicate(Num, "127.0.0.1"),
Nodes = [fake_node(N) || N <- lists:seq(1, Num)],
lists:zip3(IPs, Ports, Nodes).
fake_node(Num) ->
lists:flatten(io_lib:format("fake~p@127.0.0.1", [Num])).
add_fake_sites([Node|_], Listeners) ->
[add_site(Node, {IP, Port, fake_site(Port)})
|| {IP, Port, _} <- Listeners].
add_site(Node, {IP, Port, Name}) ->
lager:info("Add site ~p ~p:~p at node ~p", [Name, IP, Port, Node]),
Args = [IP, integer_to_list(Port), Name],
Res = rpc:call(Node, riak_repl_console, add_site, [Args]),
?assertEqual(ok, Res),
timer:sleep(timer:seconds(5)).
del_site([Node|_]=Nodes, Name) ->
lager:info("Del site ~p at ~p", [Name, Node]),
Res = rpc:call(Node, riak_repl_console, del_site, [[Name]]),
?assertEqual(ok, Res),
rt:wait_until_ring_converged(Nodes),
timer:sleep(timer:seconds(5)).
fake_site(Port) ->
lists:flatten(io_lib:format("fake_site_~p", [Port])).
verify_listeners(Listeners) ->
Strs = [IP ++ ":" ++ integer_to_list(Port) || {IP, Port, _} <- Listeners],
[?assertEqual(ok, verify_listener(Node, Strs)) || {_, _, Node} <- Listeners].
verify_listener(Node, Strs) ->
lager:info("Verify listeners ~p ~p", [Node, Strs]),
rt:wait_until(Node,
fun(_) ->
Status = rpc:call(Node, riak_repl_console, status, [quiet]),
lists:all(fun(Str) ->
lists:keymember(Str, 2, Status)
end, Strs)
end).
add_listeners(Nodes=[FirstNode|_]) ->
Ports = gen_ports(9010, length(Nodes)),
IPs = lists:duplicate(length(Nodes), "127.0.0.1"),
PN = lists:zip3(IPs, Ports, Nodes),
[add_listener(FirstNode, Node, IP, Port) || {IP, Port, Node} <- PN],
timer:sleep(timer:seconds(5)),
PN.
add_listener(N, Node, IP, Port) ->
lager:info("Adding repl listener to ~p ~s:~p", [Node, IP, Port]),
Args = [[atom_to_list(Node), IP, integer_to_list(Port)]],
Res = rpc:call(N, riak_repl_console, add_listener, Args),
?assertEqual(ok, Res).
del_listeners(Node) ->
Listeners = get_listeners(Node),
lists:foreach(fun(Listener={IP, Port, N}) ->
lager:info("deleting listener ~p on ~p", [Listener, Node]),
Res = rpc:call(Node, riak_repl_console, del_listener,
[[atom_to_list(N), IP, integer_to_list(Port)]]),
?assertEqual(ok, Res)
end, Listeners).
get_listeners(Node) ->
Status = rpc:call(Node, riak_repl_console, status, [quiet]),
%% *sigh*
[
begin
NodeName = list_to_atom(string:substr(K, 10)),
[IP, Port] = string:tokens(V, ":"),
{IP, list_to_integer(Port), NodeName}
end || {K, V} <- Status, is_list(K), string:substr(K, 1, 9) == "listener_"
].
gen_ports(Start, Len) ->
lists:seq(Start, Start + Len - 1).
verify_site_ips(Leader, Site, Listeners) ->
Status = rpc:call(Leader, riak_repl_console, status, [quiet]),
Key = lists:flatten([Site, "_ips"]),
IPStr = proplists:get_value(Key, Status),
IPs = lists:sort(re:split(IPStr, ", ")),
ExpectedIPs = lists:sort(
[list_to_binary([IP, ":", integer_to_list(Port)]) || {IP, Port, _Node} <-
Listeners]),
?assertEqual(ExpectedIPs, IPs).
make_bucket([Node|_]=Nodes, Name, Args) ->
Res = rpc:call(Node, riak_core_bucket, set_bucket, [Name, Args]),
rt:wait_until_ring_converged(Nodes),
?assertEqual(ok, Res).
start_and_wait_until_fullsync_complete(Node) ->
start_and_wait_until_fullsync_complete(Node, 20).
start_and_wait_until_fullsync_complete(Node, Retries) ->
Status0 = rpc:call(Node, riak_repl_console, status, [quiet]),
Count = proplists:get_value(server_fullsyncs, Status0) + 1,
lager:info("waiting for fullsync count to be ~p", [Count]),
lager:info("Starting fullsync on ~p (~p)", [Node,
rtdev:node_version(rtdev:node_id(Node))]),
rpc:call(Node, riak_repl_console, start_fullsync, [[]]),
%% sleep because of the old bug where stats will crash if you call it too
%% soon after starting a fullsync
timer:sleep(500),
case rt:wait_until(make_fullsync_wait_fun(Node, Count), 100, 1000) of
ok ->
ok;
_ when Retries > 0 ->
?assertEqual(ok, wait_until_connection(Node)),
lager:warning("Node failed to fullsync, retrying"),
rpc:call(Node, riak_repl_console, cancel_fullsync, [[]]),
start_and_wait_until_fullsync_complete(Node, Retries-1)
end,
lager:info("Fullsync on ~p complete", [Node]).
make_fullsync_wait_fun(Node, Count) ->
fun() ->
Status = rpc:call(Node, riak_repl_console, status, [quiet]),
case Status of
{badrpc, _} ->
false;
_ ->
case proplists:get_value(server_fullsyncs, Status) of
C when C >= Count ->
true;
_ ->
false
end
end
end.
wait_until_is_leader(Node) ->
lager:info("wait_until_is_leader(~p)", [Node]),
rt:wait_until(Node, fun is_leader/1).
is_leader(Node) ->
case rpc:call(Node, riak_repl_leader, leader_node, []) of
{badrpc, _} ->
lager:info("Badrpc"),
false;
Leader ->
lager:info("Checking: ~p =:= ~p", [Leader, Node]),
Leader =:= Node
end.
wait_until_is_not_leader(Node) ->
lager:info("wait_until_is_not_leader(~p)", [Node]),
rt:wait_until(Node, fun is_not_leader/1).
is_not_leader(Node) ->
case rpc:call(Node, riak_repl_leader, leader_node, []) of
{badrpc, _} ->
lager:info("Badrpc"),
false;
Leader ->
lager:info("Checking: ~p =/= ~p", [Leader, Node]),
Leader =/= Node
end.
wait_until_leader(Node) ->
wait_until_new_leader(Node, undefined).
wait_until_new_leader(Node, OldLeader) ->
Res = rt:wait_until(Node,
fun(_) ->
case rpc:call(Node, riak_repl_console, status, [quiet]) of
{badrpc, _} ->
false;
Status ->
case proplists:get_value(leader, Status) of
undefined ->
false;
OldLeader ->
false;
_Other ->
true
end
end
end),
?assertEqual(ok, Res).
wait_until_leader_converge([Node|_] = Nodes) ->
rt:wait_until(Node,
fun(_) ->
LeaderResults =
[get_leader(rpc:call(N, riak_repl_console, status, [quiet])) ||
N <- Nodes],
{Leaders, Errors} =
lists:partition(leader_result_filter_fun(), LeaderResults),
UniqueLeaders = lists:usort(Leaders),
Errors == [] andalso length(UniqueLeaders) == 1
end).
get_leader({badrpc, _}=Err) ->
Err;
get_leader(Status) ->
case proplists:get_value(leader, Status) of
undefined ->
false;
L ->
%%lager:info("Leader for ~p is ~p",
%%[N,L]),
L
end.
leader_result_filter_fun() ->
fun(L) ->
case L of
undefined ->
false;
{badrpc, _} ->
false;
_ ->
true
end
end.
wait_until_connection(Node) ->
rt:wait_until(Node,
fun(_) ->
case rpc:call(Node, riak_repl_console, status, [quiet]) of
{badrpc, _} ->
false;
Status ->
case proplists:get_value(server_stats, Status) of
[] ->
false;
[{_, _, too_busy}] ->
false;
[_C] ->
true;
Conns ->
lager:warning("multiple connections detected: ~p",
[Conns]),
true
end
end
end). %% 40 seconds is enough for repl
wait_until_no_connection(Node) ->
rt:wait_until(Node,
fun(_) ->
case rpc:call(Node, riak_repl_console, status, [quiet]) of
{badrpc, _} ->
false;
Status ->
case proplists:get_value(server_stats, Status) of
[] ->
true;
[{_, _, too_busy}] ->
false;
[_C] ->
false;
Conns ->
lager:warning("multiple connections detected: ~p",
[Conns]),
false
end
end
end). %% 40 seconds is enough for repl
wait_for_reads(Node, Start, End, Bucket, R) ->
rt:wait_until(Node,
fun(_) ->
rt:systest_read(Node, Start, End, Bucket, R) == []
end),
Reads = rt:systest_read(Node, Start, End, Bucket, R),
lager:info("Reads: ~p", [Reads]),
length(Reads).
do_write(Node, Start, End, Bucket, W) ->
case rt:systest_write(Node, Start, End, Bucket, W) of
[] ->
[];
Errors ->
lager:warning("~p errors while writing: ~p",
[length(Errors), Errors]),
timer:sleep(1000),
lists:flatten([rt:systest_write(Node, S, S, Bucket, W) ||
{S, _Error} <- Errors])
end.