[Bug] [StreamLoad] writing JSON has become exceptionally slow

[15767714253]

Search before asking

• I had searched in the issues and found no similar issues.

Version

2.3.5

What's Wrong?

My Table
CREATE TABLE dwd_ess_big_cell_inc
(
time datetime NOT NULL COMMENT '',
namespace_code VARCHAR(64) NOT NULL COMMENT '',
device_instance_property_code VARCHAR(64) NOT NULL COMMENT '',
device_instance_code VARCHAR(64) NOT NULL COMMENT '',
value VARCHAR(64) NULL COMMENT '',
kafka_time DATETIME NOT NULL COMMENT '创建时间',
create_time DATETIME NOT NULL DEFAULT CURRENT_TIMESTAMP COMMENT '创建时间'
) ENGINE = OLAP UNIQUE KEY( time, namespace_code,device_instance_property_code,device_instance_code)
COMMENT ''
PARTITION BY RANGE (time) ()
DISTRIBUTED BY HASH(time,namespace_code,device_instance_property_code, device_instance_code)
PROPERTIES
(
"min_load_replica_num" = "1",
"dynamic_partition.enable" = "true",
"dynamic_partition.time_unit" = "HOUR",
"dynamic_partition.start" = "-24",
"dynamic_partition.end" = "3",
"dynamic_partition.prefix" = "p",
"dynamic_partition.buckets" = "24",
"dynamic_partition.replication_num" = "3",
"compaction_policy" = "time_series",
"enable_unique_key_merge_on_write" = "false"
);

flink doris connector config

"properties": {
"format": "json",
"timezone": "Asia/Shanghai",
"read_json_by_line": "true",
"send_batch_parallelism": 10,
"memtable_on_sink_node": "true",
"columns": "time,time=from_unixtime(round(time/1000,0)),namespace_code,device_instance_property_code,device_instance_code,value,kafka_time,kafka_time=from_unixtime(round(kafka_time/1000,0))"
},

My FE Config
enable_single_replica_load = true
fetch_stream_load_record_interval_second = 30

My BE Config
number_tablet_writer_threads = 48
streaming_load_json_max_mb = 1024
enable_single_replica_load = true
jsonb_type_length_soft_limit_bytes = 2147483643
string_type_length_soft_limit_bytes = 2147483643
enable_stream_load_record = true
max_send_batch_parallelism_per_job = 20

1FE 3BE
4 * 64G 32vCpu

StreamLoad Result

Sometimes it is like this.

What You Expected?

During my previous tests, with 10 concurrent processes, each committing a million writes, it would not take more than 10 seconds. However, I don't understand why writing to the new cluster has become so slow now. This cluster only contains this single table and has plenty of resources. I hope to find out what the problem is.

How to Reproduce?

No response

Anything Else?

No response

Are you willing to submit PR?

• Yes I am willing to submit a PR!

Code of Conduct

• I agree to follow this project's Code of Conduct

[JNSimba]

Can be upgraded to 1.6.1？

[15767714253]

Can be upgraded to 1.6.1？

OK！