Merge pull request #183 from keynslug/fix/race-system-limit

fix(bootstrap): wait core tables are ready before copying
emqx · Oct 21, 2024 · 4cab8da · 4cab8da
2 parents ad1cbd7 + 6084346
commit 4cab8da
Show file tree

Hide file tree

Showing 3 changed files with 58 additions and 6 deletions.
diff --git a/src/mria_rlog.hrl b/src/mria_rlog.hrl
@@ -25,9 +25,9 @@
 -define(unexpected_event_kind, "Mria worker received unexpected event").
 -define(unexpected_event_tp(Params),
         ?tp(warning, ?unexpected_event_kind,
-            Params#{ process => ?MODULE
-                   , callback => ?FUNCTION_NAME
-                   })).
+            (begin Params end)#{ process => ?MODULE
+                               , callback => ?FUNCTION_NAME
+                               })).
 
 -define(terminate_tp,
         ?tp(debug, mria_worker_terminate, #{process => ?MODULE, callback => terminate})).

diff --git a/src/mria_schema.erl b/src/mria_schema.erl
@@ -317,19 +317,20 @@ converge_schema(Entries, InitialState) ->
 %% is needed so we can replicate schema updates just like regular
 %% transactions.
 bootstrap() ->
-    Storage = ram_copies,
     Opts = [{type, ordered_set},
             {record_name, ?schema},
             {attributes, record_info(fields, ?schema)}
            ],
     MetaSpec = #?schema{ mnesia_table = ?schema
                        , shard = ?mria_meta_shard
-                       , storage = Storage
+                       , storage = ram_copies
                        , config = Opts
                        },
     %% Create (or copy) the mnesia table and wait for it:
     ok = create_table(MetaSpec),
-    ok = mria_mnesia:copy_table(?schema, Storage),
+    %% Ensure replicas are available before starting copy:
+    ok = mria_mnesia:wait_for_tables([?schema]),
+    ok = mria_mnesia:copy_table(?schema, ram_copies),
     RlogSyncOpts = [{record_name, ?rlog_sync},
                     {attributes, record_info(fields, ?rlog_sync)}
                    ],
@@ -339,6 +340,10 @@ bootstrap() ->
                            , config = RlogSyncOpts
                            },
     ok = create_table(RlogSyncSpec),
+    %% Ensure replicas are available before starting copy:
+    %% If we've managed to sync only mnesia schema up to this point, `copy_table/2` may
+    %% fail if other nodes suddenly become unavailable.
+    ok = mria_mnesia:wait_for_tables([?rlog_sync]),
     ok = mria_mnesia:copy_table(?rlog_sync, null_copies),
     mria_mnesia:wait_for_tables([?schema, ?rlog_sync]),
     %% Seed the table with the metadata:

diff --git a/test/mria_mnesia_SUITE.erl b/test/mria_mnesia_SUITE.erl
@@ -99,6 +99,53 @@ t_join_after_node_down(_) ->
        end,
        []).
 
+%% Start a cluster of two nodes, then join the third, and simulate two nodes went down
+%% right after the third one joined. Restore them and verify the third one is healthy.
+t_cluster_down_after_join(_) ->
+    ClusterEnv = mria_mnesia_test_util:common_env(),
+    Cluster = [C1, C2, C3] = mria_ct:cluster([core, core, core], ClusterEnv),
+    ?check_trace(
+       #{timetrap => 10_000},
+       try
+           %% Prepare cluster with 3 nodes:
+           Ns = [N1, N2, N3] = mria_ct:start_cluster(node, Cluster),
+           ?assertEqual([{ok, ok} || _ <- Ns], erpc:multicall(Ns, mria, start, [])),
+           %% Join together first 2 nodes:
+           ?assertEqual(ok, erpc:call(N1, mria, join, [N2])),
+           ?assertEqual([N1, N2], lists:sort(erpc:call(N1, mria_mnesia, running_nodes, []))),
+           ?assertEqual(ok, erpc:call(N1, mria_transaction_gen, init, [])),
+           %% Tell N3 to join but simulate it goes down after joining but before bootstrap:
+           ?assertEqual(ok, erpc:call(N3, meck, new, [mria_app, [no_link, passthrough]])),
+           ?assertEqual(ok, erpc:call(N3, meck, expect, [mria_app, start, fun ?MODULE:suicide/2])),
+           %% Node N3 expectedly dies:
+           ?assertError({erpc, _}, erpc:call(N3, mria, join, [N1])),
+           ?assertError({erpc, _}, erpc:call(N3, mria_mnesia, running_nodes, [])),
+           %% Tell N1 and N2 to stop:
+           ?assertEqual(ok, erpc:call(N1, mria, stop, [])),
+           ?assertEqual(ok, erpc:call(N2, mria, stop, [])),
+           ?assertEqual([ok, ok], [slave:stop(N) || N <- [N1, N2]]),
+           %% Restart N3 and tell mria to start:
+           N3 = mria_ct:start_slave(node, C3),
+           %% This will hang waiting for N1 or N2 to go online, thus `cast/4`:
+           ?assertEqual(ok, erpc:cast(N3, mria, start, [])),
+           %% Tell N1 and N2 to get back up:
+           [N1, N2] = [mria_ct:start_slave(node, C) || C <- [C1, C2]],
+           %% Again, use `cast/4` to avoid hanging waiting for another node:
+           ?assertEqual(ok, erpc:cast(N1, mria, start, [])),
+           ?assertEqual(ok, erpc:cast(N2, mria, start, [])),
+           ?assertEqual(ok, erpc:call(N3, mria, start, [])),
+           %% Verify that bootstrap process has finished and the node is alive:
+           _ = erpc:call(N3, sys, get_state, [mria_schema]),
+           ?assertEqual([N1, N2, N3], lists:sort(erpc:call(N3, mria_mnesia, running_nodes, []))),
+           ok
+       after
+           ok = mria_ct:teardown_cluster(Cluster)
+       end,
+       []).
+
+suicide(_Type, _Args) ->
+    erlang:halt().
+
 t_diagnosis_tab(_)->
     TestTab = test_tab_1,
     Cluster = [NS1, NS2] = mria_ct:cluster([core, core], []),