From 43f80e263377d07429fb082ed9ade746b4d910bd Mon Sep 17 00:00:00 2001 From: Xin Liao Date: Fri, 11 Nov 2022 18:05:36 +0800 Subject: [PATCH] [enhancement](load) Increase batch size of node channel to improve import performance (#13912) --- be/src/vec/sink/vtablet_sink.cpp | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/be/src/vec/sink/vtablet_sink.cpp b/be/src/vec/sink/vtablet_sink.cpp index a6e4a96566..3599abcbb3 100644 --- a/be/src/vec/sink/vtablet_sink.cpp +++ b/be/src/vec/sink/vtablet_sink.cpp @@ -68,6 +68,12 @@ Status VNodeChannel::init(RuntimeState* state) { _cur_add_block_request.set_eos(false); _name = fmt::format("VNodeChannel[{}-{}]", _index_channel->_index_id, _node_id); + // The node channel will send _batch_size rows of data each rpc. When the + // number of tablets is large, the number of data rows received by each + // tablet is small, TabletsChannel need to traverse each tablet for import. + // so the import performance is poor. Therefore, we set _batch_size to + // a relatively large value to improve the import performance. + _batch_size = std::max(_batch_size, 8192); return Status::OK(); }