Zhong Wang created IOTDB-2042:
---------------------------------
Summary: Data loss after restart
Key: IOTDB-2042
URL: https://issues.apache.org/jira/browse/IOTDB-2042
Project: Apache IoTDB
Issue Type: Bug
Reporter: Zhong Wang
Way to reproduce:
1. Enable time partitioning. Set the partition_interval to 2592000 (30 days).
2. Start a 3 node cluster with 2 replicas.
3. Insert the following values on node 3:
{code:sql}
insert into root.group_1.d_1(timestamp, s_1)
values(2018-07-18T00:00:00.000+08:00, 18.0)
insert into root.group_1.d_1(timestamp, s_1)
values(2018-07-19T00:00:00.000+08:00, 19.0)
insert into root.group_1.d_1(timestamp, s_1)
values(2019-08-19T00:00:00.000+08:00, 20.0)
{code}
4. Restart the cluster.
5. Execute the following query on each node. Node 2 will return an empty result
set:
{code:sql}
SELECT * FROM root
{code}
Some of the TSFiles are empty after recovery, which is unexpected.
Reason:
When time partitioning is enabled, although the data of a time series can be
distributed across the whole cluster, its metadata is only stored in the data
group corresponding to the hash value calculated from its storage group and
timestamp 0.
When a node from another data group receives insertions of the time series, it
will send a request to whoever has the metadata and cache the response, but the
cache will not be persisted. Thus, if the node crashes, it is not able to
recover from the WAL since the cache is not available anymore.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
