[tarantool-patches] Re: [PATCH] [replication] make join stage more informative

  • From: Vladimir Davydov <vdavydov.dev@xxxxxxxxx>
  • To: Serge Petrenko <sergepetrenko@xxxxxxxxxxxxx>
  • Date: Thu, 18 Oct 2018 15:43:13 +0300

On Tue, Oct 16, 2018 at 01:10:53PM +0300, Serge Petrenko wrote:

This patch adds logging amount of rows recieved by applier during the
join stage, the same way that recovery has it. It also adds

    downstream:
      status: joining

message to master's box.info.replication for a replica that is joining.

Please add a test case for the new box.info.replication status.


Closes #3165
---
https://github.com/tarantool/tarantool/issues/3165
https://github.com/tarantool/tarantool/tree/sp/gh-3165-join-progress-report

 src/box/applier.cc | 7 +++++++
 src/box/lua/info.c | 9 +++++++--
 2 files changed, 14 insertions(+), 2 deletions(-)

diff --git a/src/box/applier.cc b/src/box/applier.cc
index 7da278e68..ba68c3d07 100644
--- a/src/box/applier.cc
+++ b/src/box/applier.cc
@@ -309,11 +309,15 @@ applier_join(struct applier *applier)
       * Receive initial data.
       */
      assert(applier->join_stream != NULL);
+     uint64_t row_count = 0;
      while (true) {
              coio_read_xrow(coio, ibuf, &row);
              applier->last_row_time = ev_monotonic_now(loop());
              if (iproto_type_is_dml(row.type)) {
                      xstream_write_xc(applier->join_stream, &row);
+                     ++row_count;
+                     if (row_count % 100000 == 0)
+                             say_info("%.1fM rows recieved", row_count / 
1e6);

Without knowing the total number of rows to be received, this isn't very
informative IMO.

              } else if (row.type == IPROTO_OK) {
                      if (applier->version_id < version_id(1, 7, 0)) {
                              /*
@@ -354,6 +358,9 @@ applier_join(struct applier *applier)
              if (iproto_type_is_dml(row.type)) {
                      vclock_follow_xrow(&replicaset.vclock, &row);
                      xstream_write_xc(applier->subscribe_stream, &row);
+                     ++row_count;
+                     if (row_count % 100000 == 0)
+                             say_info("%.1fM rows recieved", row_count / 
1e6);

When bootstrapping from Tarantool < 1.7, final join won't end here, it
will end in applier_subscribe(). Not sure if we want to handle this case
though.

              } else if (row.type == IPROTO_OK) {
                      /*
                       * Current vclock. This is not used now,
diff --git a/src/box/lua/info.c b/src/box/lua/info.c
index 655768ec4..4c25255c0 100644
--- a/src/box/lua/info.c
+++ b/src/box/lua/info.c
@@ -119,8 +119,13 @@ static void
 lbox_pushrelay(lua_State *L, struct relay *relay)
 {
      lua_newtable(L);
-     lua_pushstring(L, "vclock");
-     lbox_pushvclock(L, relay_vclock(relay));
+     if (relay_vclock(relay)->map == 0) {
+             lua_pushstring(L, "status");
+             lua_pushstring(L, "joining");
+     } else {
+             lua_pushstring(L, "vclock");
+             lbox_pushvclock(L, relay_vclock(relay));
+     }

This wouldn't work, because struct replica doesn't exist before initial
join is complete.

      lua_settable(L, -3);
 }

Other related posts:

  1. Home
  2. tarantool-patches
  3. Archive
  4. 10-2018