<div dir="ltr"><div dir="ltr"><div>Hi, Vlad</div><div><br></div><div>Thanks for the fix a lot. In general it looks good to me, but I've got few questions. Find them below.<br></div><div><br></div><div><div><div><div dir="ltr" data-smartmail="gmail_signature"><div dir="ltr"><span><div><div dir="ltr">Best regards</div><div>Yaroslav Dynnikov<br></div></div></span></div></div></div><br></div></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, 5 Aug 2020 at 00:04, Vladislav Shpilevoy <<a href="mailto:v.shpilevoy@tarantool.org" target="_blank">v.shpilevoy@tarantool.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Before the patch a replica couldn't call vshard.storage.cfg()<br>
before the master is bootstrapped. That could lead to an exception<br>
in the middle of vshard.storage.cfg().<br>
<br>
It sounds not so bad and even looks unreachable, because replica<br>
bootstrap anyway is blocked in box.cfg() until a master node is<br>
started, otherwise the replica instance is terminated.<br>
But the box.cfg bootstrap != vshard bootstrap.<br>
<br>
It could be that both replica and master nodes are already<br>
bootstrapped and working, but vshard.storage.cfg() wasn't called<br>
yet. In that case vshard.storage.cfg() on the replica wouldn't<br>
block in box.cfg(), and would happily fail a few lines later in<br>
vshard.storage.cfg() at attempt to touch not existing<br>
box.space._bucket.<br>
<br>
That was the case for cartridge. The framework configures<br>
instances in its own way before vshard.storage.cfg(). Then it<br>
calls vshard.storage.cfg() on them in arbitrary order, and<br>
sometimes replicas were configured earlier than the master.<br>
<br>
The fail is fixed by simply skipping the _bucket's trigger<br>
installation on the replica node. Because so far it is not needed<br>
here for anything. The trigger's sole purpose is to increment<br>
bucket generation used only for DML on _bucket on the master node.<br>
<br>
Now vshard.storage.cfg() on replica is able to finish before<br>
master does the same. However the replica still won't be able to<br>
handle vshard.storage.* requests until receives vshard schema<br>
from master.<br>
<br>
Closes #237<br>
---<br>
Branch: <a href="http://github.com/tarantool/vshard/tree/gerold103/gh-237-boot-replica-first" rel="noreferrer" target="_blank">http://github.com/tarantool/vshard/tree/gerold103/gh-237-boot-replica-first</a><br>
Issue: <a href="https://github.com/tarantool/vshard/issues/237" rel="noreferrer" target="_blank">https://github.com/tarantool/vshard/issues/237</a><br>
<br>
test/lua_libs/storage_template.lua | 55 +++++++++++-<br>
test/reload_evolution/storage.result | 10 +++<br>
test/reload_evolution/storage.test.lua | 8 ++<br>
test/router/boot_replica_first.result | 112 ++++++++++++++++++++++++<br>
test/router/boot_replica_first.test.lua | 42 +++++++++<br>
vshard/storage/init.lua | 18 +++-<br>
vshard/storage/reload_evolution.lua | 10 ++-<br>
7 files changed, 251 insertions(+), 4 deletions(-)<br>
create mode 100644 test/router/boot_replica_first.result<br>
create mode 100644 test/router/boot_replica_first.test.lua<br>
<br>
diff --git a/test/lua_libs/storage_template.lua b/test/lua_libs/storage_template.lua<br>
index 29a753d..84e4180 100644<br>
--- a/test/lua_libs/storage_template.lua<br>
+++ b/test/lua_libs/storage_template.lua<br>
@@ -1,5 +1,7 @@<br>
#!/usr/bin/env tarantool<br>
<br>
+local luri = require('uri')<br>
+<br>
NAME = require('fio').basename(arg[0], '.lua')<br>
fiber = require('fiber')<br>
test_run = require('test_run').new()<br>
@@ -10,12 +12,63 @@ log = require('log')<br>
if not cfg.shard_index then<br>
cfg.shard_index = 'bucket_id'<br>
end<br>
+instance_uuid = util.name_to_uuid[NAME]<br>
+<br>
+--<br>
+-- Bootstrap the instance exactly like vshard does. But don't<br>
+-- initialize any vshard-specific code.<br>
+--<br>
+local function boot_like_vshard()<br>
+ assert(type(box.cfg) == 'function')<br>
+ for rs_uuid, rs in pairs(cfg.sharding) do<br>
+ for replica_uuid, replica in pairs(rs.replicas) do<br>
+ if replica_uuid == instance_uuid then<br>
+ local box_cfg = {replication = {}}<br>
+ box_cfg.instance_uuid = replica_uuid<br>
+ box_cfg.replicaset_uuid = rs_uuid<br>
+ box_cfg.listen = replica.uri<br>
+ box_cfg.read_only = not replica.master<br>
+ box_cfg.replication_connect_quorum = 0<br>
+ box_cfg.replication_timeout = 0.1<br>
+ for _, replica in pairs(rs.replicas) do<br>
+ table.insert(box_cfg.replication, replica.uri)<br>
+ end<br>
+ box.cfg(box_cfg)<br>
+ if not replica.master then<br>
+ return<br>
+ end<br>
+ local uri = luri.parse(replica.uri)<br>
+ box.schema.user.create(uri.login, {<br>
+ password = uri.password, if_not_exists = true,<br>
+ })<br>
+ box.schema.user.grant(uri.login, 'super')<br>
+ return<br>
+ end<br>
+ end<br>
+ end<br>
+ assert(false)<br></blockquote><div><br></div><div>Should there be some meaningful message?<br></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
+end<br>
+<br>
+local omit_cfg = false<br>
+local i = 1<br>
+while arg[i] ~= nil do<br>
+ local key = arg[i]<br>
+ i = i + 1<br>
+ if key == 'boot_before_cfg' then<br>
+ boot_like_vshard()<br>
+ omit_cfg = true<br>
+ end<br>
+end<br>
<br>
vshard = require('vshard')<br>
echo_count = 0<br>
cfg.replication_connect_timeout = 3<br>
cfg.replication_timeout = 0.1<br>
-vshard.storage.cfg(cfg, util.name_to_uuid[NAME])<br>
+<br>
+if not omit_cfg then<br>
+ vshard.storage.cfg(cfg, instance_uuid)<br>
+end<br>
+<br>
function bootstrap_storage(engine)<br>
box.once("testapp:schema:1", function()<br>
if rawget(_G, 'CHANGE_SPACE_IDS') then<br>
diff --git a/test/reload_evolution/storage.result b/test/reload_evolution/storage.result<br>
index ebf4fdc..4652c4f 100644<br>
--- a/test/reload_evolution/storage.result<br>
+++ b/test/reload_evolution/storage.result<br>
@@ -92,6 +92,16 @@ test_run:grep_log('storage_2_a', 'vshard.storage.reload_evolution: upgraded to')<br>
...<br>
vshard.storage.internal.reload_version<br>
---<br>
+- 2<br>
+...<br>
+--<br>
+-- gh-237: should be only one trigger. During gh-237 the trigger installation<br>
+-- became conditional and therefore required to remember the current trigger<br>
+-- somewhere. When reloaded from the old version, the trigger needed to be<br>
+-- fetched from _bucket:on_replace().<br>
+--<br>
+#box.space._bucket:on_replace()<br>
+---<br>
- 1<br>
...<br>
-- Make sure storage operates well.<br>
diff --git a/test/reload_evolution/storage.test.lua b/test/reload_evolution/storage.test.lua<br>
index 56c1693..06f7117 100644<br>
--- a/test/reload_evolution/storage.test.lua<br>
+++ b/test/reload_evolution/storage.test.lua<br>
@@ -37,6 +37,14 @@ package.loaded['vshard.storage'] = nil<br>
vshard.storage = require("vshard.storage")<br>
test_run:grep_log('storage_2_a', 'vshard.storage.reload_evolution: upgraded to') ~= nil<br>
vshard.storage.internal.reload_version<br>
+--<br>
+-- gh-237: should be only one trigger. During gh-237 the trigger installation<br>
+-- became conditional and therefore required to remember the current trigger<br>
+-- somewhere. When reloaded from the old version, the trigger needed to be<br>
+-- fetched from _bucket:on_replace().<br>
+--<br>
+#box.space._bucket:on_replace()<br>
+<br>
-- Make sure storage operates well.<br>
vshard.storage.bucket_force_drop(2000)<br>
vshard.storage.bucket_force_create(2000)<br>
diff --git a/test/router/boot_replica_first.result b/test/router/boot_replica_first.result<br>
new file mode 100644<br>
index 0000000..1705230<br>
--- /dev/null<br>
+++ b/test/router/boot_replica_first.result<br>
@@ -0,0 +1,112 @@<br>
+-- test-run result file version 2<br>
+test_run = require('test_run').new()<br>
+ | ---<br>
+ | ...<br>
+REPLICASET_1 = { 'box_1_a', 'box_1_b', 'box_1_c' }<br>
+ | ---<br>
+ | ...<br>
+test_run:create_cluster(REPLICASET_1, 'router', {args = 'boot_before_cfg'})<br>
+ | ---<br>
+ | ...<br>
+util = require('util')<br>
+ | ---<br>
+ | ...<br>
+util.wait_master(test_run, REPLICASET_1, 'box_1_a')<br>
+ | ---<br>
+ | ...<br>
+_ = test_run:cmd("create server router with script='router/router_2.lua'")<br>
+ | ---<br>
+ | ...<br>
+_ = test_run:cmd("start server router")<br>
+ | ---<br>
+ | ...<br>
+<br>
+--<br>
+-- gh-237: replica should be able to boot before master. Before the issue was<br>
+-- fixed, replica always tried to install a trigger on _bucket space even when<br>
+-- it was not created on a master yet - that lead to an exception in<br>
+-- storage.cfg. Now it should not install the trigger at all, because anyway it<br>
+-- is not needed on replica for anything.<br>
+--<br>
+<br>
+test_run:switch('box_1_b')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+vshard.storage.cfg(cfg, instance_uuid)<br>
+ | ---<br>
+ | ...<br>
+-- _bucket is not created yet. Will fail.<br>
+util.check_error(vshard.storage.call, 1, 'read', 'echo', {100})<br>
+ | ---<br>
+ | - attempt to index field '_bucket' (a nil value)<br>
+ | ...<br>
+<br>
+test_run:switch('default')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+util.map_evals(test_run, {REPLICASET_1}, 'bootstrap_storage(\'memtx\')')<br>
+ | ---<br>
+ | ...<br>
+<br>
+test_run:switch('box_1_a')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+vshard.storage.cfg(cfg, instance_uuid)<br>
+ | ---<br>
+ | ...<br>
+<br>
+test_run:switch('box_1_b')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+test_run:wait_lsn('box_1_b', 'box_1_a')<br>
+ | ---<br>
+ | ...<br>
+-- Fails, but gracefully. Because the bucket is not found here.<br>
+vshard.storage.call(1, 'read', 'echo', {100})<br>
+ | ---<br>
+ | - null<br>
+ | - bucket_id: 1<br>
+ | reason: Not found<br>
+ | code: 1<br>
+ | type: ShardingError<br>
+ | message: 'Cannot perform action with bucket 1, reason: Not found'<br>
+ | name: WRONG_BUCKET<br>
+ | ...<br>
+-- Should not have triggers.<br>
+#box.space._bucket:on_replace()<br>
+ | ---<br>
+ | - 0<br>
+ | ...<br>
+<br>
+test_run:switch('router')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+vshard.router.bootstrap()<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+vshard.router.callro(1, 'echo', {100})<br>
+ | ---<br>
+ | - 100<br>
+ | ...<br>
+<br>
+test_run:switch("default")<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+test_run:cmd('stop server router')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+test_run:cmd('delete server router')<br>
+ | ---<br>
+ | - true<br>
+ | ...<br>
+test_run:drop_cluster(REPLICASET_1)<br>
+ | ---<br>
+ | ...<br>
diff --git a/test/router/boot_replica_first.test.lua b/test/router/boot_replica_first.test.lua<br>
new file mode 100644<br>
index 0000000..7b1b3fd<br>
--- /dev/null<br>
+++ b/test/router/boot_replica_first.test.lua<br>
@@ -0,0 +1,42 @@<br>
+test_run = require('test_run').new()<br>
+REPLICASET_1 = { 'box_1_a', 'box_1_b', 'box_1_c' }<br>
+test_run:create_cluster(REPLICASET_1, 'router', {args = 'boot_before_cfg'})<br>
+util = require('util')<br>
+util.wait_master(test_run, REPLICASET_1, 'box_1_a')<br>
+_ = test_run:cmd("create server router with script='router/router_2.lua'")<br>
+_ = test_run:cmd("start server router")<br>
+<br>
+--<br>
+-- gh-237: replica should be able to boot before master. Before the issue was<br>
+-- fixed, replica always tried to install a trigger on _bucket space even when<br>
+-- it was not created on a master yet - that lead to an exception in<br>
+-- storage.cfg. Now it should not install the trigger at all, because anyway it<br>
+-- is not needed on replica for anything.<br>
+--<br>
+<br>
+test_run:switch('box_1_b')<br>
+vshard.storage.cfg(cfg, instance_uuid)<br>
+-- _bucket is not created yet. Will fail.<br>
+util.check_error(vshard.storage.call, 1, 'read', 'echo', {100})<br>
+<br>
+test_run:switch('default')<br>
+util.map_evals(test_run, {REPLICASET_1}, 'bootstrap_storage(\'memtx\')')<br>
+<br>
+test_run:switch('box_1_a')<br>
+vshard.storage.cfg(cfg, instance_uuid)<br>
+<br>
+test_run:switch('box_1_b')<br>
+test_run:wait_lsn('box_1_b', 'box_1_a')<br>
+-- Fails, but gracefully. Because the bucket is not found here.<br>
+vshard.storage.call(1, 'read', 'echo', {100})<br>
+-- Should not have triggers.<br>
+#box.space._bucket:on_replace()<br>
+<br>
+test_run:switch('router')<br>
+vshard.router.bootstrap()<br>
+vshard.router.callro(1, 'echo', {100})<br>
+<br>
+test_run:switch("default")<br>
+test_run:cmd('stop server router')<br>
+test_run:cmd('delete server router')<br>
+test_run:drop_cluster(REPLICASET_1)<br>
diff --git a/vshard/storage/init.lua b/vshard/storage/init.lua<br>
index c6a78fe..ed577f9 100644<br>
--- a/vshard/storage/init.lua<br>
+++ b/vshard/storage/init.lua<br>
@@ -94,6 +94,17 @@ if not M then<br>
-- detect that _bucket was not changed between yields.<br>
--<br>
bucket_generation = 0,<br>
+ --<br>
+ -- Reference to the function used as on_replace trigger on<br>
+ -- _bucket space. It is used to replace the trigger with<br>
+ -- a new function when reload happens. It is kept<br>
+ -- explicitly because the old function is deleted on<br>
+ -- reload from the global namespace. On the other hand, it<br>
+ -- is still stored in _bucket:on_replace() somewhere, but<br>
+ -- it is not known where. The only 100% way to be able to<br>
+ -- replace the old function is to keep its reference.<br>
+ --<br>
+ bucket_on_replace = nil,<br>
<br>
------------------- Garbage collection -------------------<br>
-- Fiber to remove garbage buckets data.<br>
@@ -2435,8 +2446,11 @@ local function storage_cfg(cfg, this_replica_uuid, is_reload)<br>
local uri = luri.parse(this_replica.uri)<br>
schema_upgrade(is_master, uri.login, uri.password)<br>
<br>
- local old_trigger = box.space._bucket:on_replace()[1]<br>
- box.space._bucket:on_replace(bucket_generation_increment, old_trigger)<br>
+ if is_master then<br>
+ local old_trigger = M.bucket_on_replace<br>
+ box.space._bucket:on_replace(bucket_generation_increment, old_trigger)<br>
+ M.bucket_on_replace = bucket_generation_increment<br>
+ end<br>
<br>
lreplicaset.rebind_replicasets(new_replicasets, M.replicasets)<br>
lreplicaset.outdate_replicasets(M.replicasets)<br>
diff --git a/vshard/storage/reload_evolution.lua b/vshard/storage/reload_evolution.lua<br>
index 5d09b11..f38af74 100644<br>
--- a/vshard/storage/reload_evolution.lua<br>
+++ b/vshard/storage/reload_evolution.lua<br>
@@ -17,6 +17,14 @@ migrations[#migrations + 1] = function(M)<br>
-- Code to update Lua objects.<br>
end<br>
<br>
+migrations[#migrations + 1] = function(M)<br>
+ local bucket = box.space._bucket<br>
+ if bucket then<br>
+ assert(M.bucket_on_replace == nil)<br>
+ M.bucket_on_replace = bucket:on_replace()[1]<br>
+ end<br>
+end<br>
+<br>
--<br>
-- Perform an update based on a version stored in `M` (internals).<br>
-- @param M Old module internals which should be updated.<br>
@@ -33,7 +41,7 @@ local function upgrade(M)<br>
log.error(err_msg)<br>
error(err_msg)<br>
end<br>
- for i = start_version, #migrations do<br>
+ for i = start_version + 1, #migrations do<br></blockquote><div><br></div><div>Is it already tested somewhere else (migrations I mean)? This change looks like a fix for some other problem.<br></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
local ok, err = pcall(migrations[i], M)<br>
if ok then<br>
<a href="http://log.info" rel="noreferrer" target="_blank">log.info</a>('vshard.storage.reload_evolution: upgraded to %d version',<br>
-- <br>
2.21.1 (Apple Git-122.3)<br>
<br>
</blockquote></div></div>