storage: return to fullmesh topology

Fullmesh was disabled by #66 because of its too raw
implementation in Tarantool. Now it is ripe enough to be
granted a second chance.

Fullmesh provides a benefit of a constant complete sync. If a
tuple is written on a master right before its demotion, a new
master still will receive the tuple from the old one.

Closes #83

test/misc/check_uuid_on_connect.result

@@ -98,15 +98,15 @@ test_run:cmd('create server bad_uuid_2_a_repaired with script="misc/bad_uuid_2_a
 ---
 - true
 ...
-test_run:cmd('start server bad_uuid_2_a_repaired')
+test_run:cmd('start server bad_uuid_2_a_repaired with wait_load=False, wait=False')
 ---
 - true
 ...
 test_run:cmd('create server bad_uuid_2_b with script="misc/bad_uuid_2_b.lua", wait=False, wait_load=False')
 ---
 - true
 ...
-test_run:cmd('start server bad_uuid_2_b')
+test_run:cmd('start server bad_uuid_2_b with wait_load=False, wait=False')
 ---
 - true
 ...

test/misc/check_uuid_on_connect.test.lua

@@ -37,9 +37,9 @@ test_run:cmd("switch default")
 test_run:drop_cluster(REPLICASET_2)
 REPLICASET_2 = { 'bad_uuid_2_a_repaired', 'bad_uuid_2_b' }
 test_run:cmd('create server bad_uuid_2_a_repaired with script="misc/bad_uuid_2_a_repaired.lua", wait=False, wait_load=False')
-test_run:cmd('start server bad_uuid_2_a_repaired')
+test_run:cmd('start server bad_uuid_2_a_repaired with wait_load=False, wait=False')
 test_run:cmd('create server bad_uuid_2_b with script="misc/bad_uuid_2_b.lua", wait=False, wait_load=False')
-test_run:cmd('start server bad_uuid_2_b')
+test_run:cmd('start server bad_uuid_2_b with wait_load=False, wait=False')
 util.wait_master(test_run, REPLICASET_2, 'bad_uuid_2_a_repaired')
 
 test_run:switch('bad_uuid_1_a')

test/misc/fullmesh.result

@@ -0,0 +1,99 @@
+test_run = require('test_run').new()
+---
+...
+REPLICASET_1 = { 'storage_1_a', 'storage_1_b' }
+---
+...
+REPLICASET_2 = { 'storage_2_a', 'storage_2_b' }
+---
+...
+test_run:create_cluster(REPLICASET_1, 'misc')
+---
+...
+test_run:create_cluster(REPLICASET_2, 'misc')
+---
+...
+util = require('util')
+---
+...
+test_run:wait_fullmesh(REPLICASET_1)
+---
+...
+test_run:wait_fullmesh(REPLICASET_2)
+---
+...
+util.map_evals(test_run, {REPLICASET_1, REPLICASET_2}, 'bootstrap_storage(\'memtx\')')
+---
+...
+--
+-- gh-83: use fullmesh topology in vshard. Scenario of the test:
+-- start writing a tuple on a master. Then switch master. After
+-- switch the tuple still should be sent from the slave to a new
+-- master.
+--
+_ = test_run:switch('storage_1_a')
+---
+...
+-- Block new requests sending.
+box.error.injection.set("ERRINJ_WAL_DELAY", true)
+---
+- ok
+...
+f = fiber.create(function() box.space.test:replace{1, 1} end)
+---
+...
+box.space.test:select{}
+---
+- - [1, 1]
+...
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_b].master = true
+---
+...
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_a].master = false
+---
+...
+vshard.storage.cfg(cfg, util.name_to_uuid.storage_1_a)
+---
+...
+_ = test_run:switch('storage_1_b')
+---
+...
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_b].master = true
+---
+...
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_a].master = false
+---
+...
+vshard.storage.cfg(cfg, util.name_to_uuid.storage_1_b)
+---
+...
+box.space.test:select{}
+---
+- []
+...
+_ = test_run:switch('storage_1_a')
+---
+...
+box.error.injection.set("ERRINJ_WAL_DELAY", false)
+---
+- ok
+...
+_ = test_run:switch('storage_1_b')
+---
+...
+while box.space.test:count() == 0 do fiber.sleep(0.1) end
+---
+...
+box.space.test:select{}
+---
+- - [1, 1]
+...
+_ = test_run:cmd("switch default")
+---
+...
+test_run:drop_cluster(REPLICASET_2)
+---
+...
+test_run:drop_cluster(REPLICASET_1)
+---
+...

test/misc/fullmesh.test.lua

@@ -0,0 +1,39 @@
+test_run = require('test_run').new()
+REPLICASET_1 = { 'storage_1_a', 'storage_1_b' }
+REPLICASET_2 = { 'storage_2_a', 'storage_2_b' }
+test_run:create_cluster(REPLICASET_1, 'misc')
+test_run:create_cluster(REPLICASET_2, 'misc')
+util = require('util')
+test_run:wait_fullmesh(REPLICASET_1)
+test_run:wait_fullmesh(REPLICASET_2)
+util.map_evals(test_run, {REPLICASET_1, REPLICASET_2}, 'bootstrap_storage(\'memtx\')')
+
+--
+-- gh-83: use fullmesh topology in vshard. Scenario of the test:
+-- start writing a tuple on a master. Then switch master. After
+-- switch the tuple still should be sent from the slave to a new
+-- master.
+--
+
+_ = test_run:switch('storage_1_a')
+-- Block new requests sending.
+box.error.injection.set("ERRINJ_WAL_DELAY", true)
+f = fiber.create(function() box.space.test:replace{1, 1} end)
+box.space.test:select{}
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_b].master = true
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_a].master = false
+vshard.storage.cfg(cfg, util.name_to_uuid.storage_1_a)
+_ = test_run:switch('storage_1_b')
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_b].master = true
+cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_a].master = false
+vshard.storage.cfg(cfg, util.name_to_uuid.storage_1_b)
+box.space.test:select{}
+_ = test_run:switch('storage_1_a')
+box.error.injection.set("ERRINJ_WAL_DELAY", false)
+_ = test_run:switch('storage_1_b')
+while box.space.test:count() == 0 do fiber.sleep(0.1) end
+box.space.test:select{}
+
+_ = test_run:cmd("switch default")
+test_run:drop_cluster(REPLICASET_2)
+test_run:drop_cluster(REPLICASET_1)

test/misc/master_switch.result

@@ -57,6 +57,9 @@ cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_b].master
 cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_a].master = false
 ---
 ...
+cfg.replication_connect_quorum = 1
+---
+...
 vshard.storage.cfg(cfg, util.name_to_uuid.storage_1_b)
 ---
 ...

test/misc/master_switch.test.lua

@@ -19,6 +19,7 @@ _ = test_run:cmd('start server storage_1_b')
 _ = test_run:switch('storage_1_b')
 cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_b].master = true
 cfg.sharding[util.replicasets[1]].replicas[util.name_to_uuid.storage_1_a].master = false
+cfg.replication_connect_quorum = 1
 vshard.storage.cfg(cfg, util.name_to_uuid.storage_1_b)
 box.space._bucket:replace({1, vshard.consts.BUCKET.ACTIVE})
 box.space.test:insert{1, 1, 2}

test/misc/reconfigure.result

@@ -275,7 +275,7 @@ uris
 ...
 box.cfg.replication
 ---
-- []
+- - storage:[email protected]:3301
 ...
 _ = test_run:switch('storage_2_a')
 ---
@@ -300,7 +300,8 @@ uris
 ...
 box.cfg.replication
 ---
-- - storage:[email protected]:3304
+- - storage:[email protected]:3303
+  - storage:[email protected]:3304
 ...
 _ = test_run:switch('storage_2_b')
 ---
@@ -325,7 +326,8 @@ uris
 ...
 box.cfg.replication
 ---
-- []
+- - storage:[email protected]:3303
+  - storage:[email protected]:3304
 ...
 _ = test_run:switch('storage_3_a')
 ---
@@ -350,7 +352,7 @@ uris
 ...
 box.cfg.replication
 ---
-- []
+- - storage:[email protected]:3306
 ...
 _ = test_run:switch('router_1')
 ---

test/misc/suite.ini

@@ -3,4 +3,5 @@ core = tarantool
 description = Misc tests
 script = test.lua
 is_parallel = False
+release_disabled = fullmesh.test.lua
 lua_libs = ../lua_libs/util.lua bad_uuid_config.lua ../../example/localcfg.lua

vshard/storage/init.lua

@@ -1791,9 +1791,11 @@ local function storage_cfg(cfg, this_replica_uuid, is_reload)
         -- If a master role of the replica is not changed, then
         -- 'read_only' can be set right here.
         box_cfg.listen = box_cfg.listen or this_replica.uri
-        if box_cfg.replication == nil and this_replicaset.master
-           and not is_master then
-            box_cfg.replication = {this_replicaset.master.uri}
+        if not box_cfg.replication then
+            box_cfg.replication = {}
+            for uuid, replica in pairs(this_replicaset.replicas) do
+                table.insert(box_cfg.replication, replica.uri)
+            end
         else
             box_cfg.replication = {}
         end