...
TODO: need more detail design.
4. Single remote replica mode
Description
The current cold and hot data is divided into two parts. The hot data is stored on the local disk and the cold data is stored in the remote cluster. Due to the independence of segments between BEs, cold data has multiple copies on the remote cluster, resulting in data redundancy. To solve this problem, the single remote replica is proposed.
A single remote replica, as the name implies, means that the number of data copies saved on the remote cluster is one, that is, no additional layer of copy management is added to the original file system copy management of the cluster, so as to greatly reduce the storage space occupied.
In order to meet the cooperative management between different replicas, the FE selects one of several replicas as the upload node, while the replica that is not selected will only check whether the current segment data has been uploaded on the cluster, and synchronously delete the local data. The copy of the selected upload file will upload the data file normally, and upload the data of the current tablet to the cluster for other fragment checks.
OrganizationChart
CooldownHandler in FE
CooldownHandler uniformly manages the configuration of data transfer to cold, which only takes effect in the MASTER node. There is a list that records all the tablets configured with cold and hot data rules and their replicas, and checks the status of each replica. The processing here is only for ordinary tablets and does not operate on ShadowTablets. The replica information is consistent with the show tablet.
When the CooldownType of a tablet is inconsistent with the status reported by BE, the CooldownType status needs to be synchronized first (first turn off the upload of the upload copy of the file being uploaded and set it to USE_REMOTE_DATA; then set the new upload copy status to UPLOAD_DATA). When the copy of a tablet is abnormal (the CooldownType of all copies is USE_REMOTE_DATA, the tablet with the CooldownType specified as UPLOAD_DATA is abnormal, and the tablet with the CooldownType specified as UPLOAD_DATA has not reported the heartbeat (lastUpdateTime) after the timeout period, it is determined that the tablet needs to reselect the Upload copy and set the CooldownType to UPLOAD_ DATA。
The CooldownType on the FE side can only be accessed from the USE_ REMOTE_ DATA changes to UPLOAD_ DATA，UPLOAD_ Generally, the DATA node does not need to be changed unless BE times out. When BE times out, the FE side will automatically process the TABLET migration. The newly migrated TABLET is defaulted to USE_ REMOTE_ DATA, UPLOAD will be re specified_ DATA node. Ensure that two CooldownTypes are not UPLOAD at the same time_ Node of DATA.
Considering that the coolown operation does not require high real-time performance, and in order to prevent UPLOAD_ DATA switches frequently, and the timeout threshold can be set longer, such as half an hour.
Cooldown in BE
The CooldownType status of each tablet on the BE side is USE by default_ REMOTE_ DATA, that is, do not upload. Only when FE sends a request to set CooldownType to UPLOAD_ Only DATA updates its status and sets it to upload.
If the heartbeat time reported by the Tablet last exceeds the timeout threshold mentioned above, according to the timeout processing mechanism of the FE end, the current Tablet has been set to USE_ REMOTE_ The possibility of DATA. At this time, set CooldownType to USE_ REMOTE_ DATA, stop uploading. Later, wait for FE to re initiate the request and redistribute CooldownType to BE.
In the process of uploading files on the BE side, if the operation takes too long, the timeout threshold should also be checked regularly. If timeout is found, the current upload request should be stopped immediately and the uploaded files should be deleted. (If the remote TabletMeta has been updated, it indicates that the operation has been completed, and the file does not need to be deleted)
UPLOAD_ DATA is a super operation with high risk. It is necessary to ensure that only one copy is performing upload requests at the same time. Therefore, some real-time performance can be sacrificed to reduce the probability of this risk.
BE in UPLOAD_ DATA type
Because different copies of the same batch of imported data under the same tablet rowset_ The IDs are different. When uploading a segment, its corresponding TabletMeta should be uploaded for reference by other copies of the tablet. This information only needs to include the Upload segment, not the local segment.
When in coolown, first read the TabletMeta file under the remote cluster directory, check the existing data range (which segments have been uploaded) on the remote cluster according to it, and compare it with the local TabletMeta. If the remote TabletMeta of the local record is different, merge the remote TabletMeta.
According to CooldownType status, there are two operations:
If CooldownType is UPLOAD_ DATA, indicating that the current tablet needs to be uploaded. Upload the subsequent segments according to the data range on the cluster, and then update the remote TabletMeta. During the upload process, the CooldownType should also be checked regularly to prevent the CooldownType update from not being found because the upload time is too long.
If CooldownType is USE_ REMOTE_ DATA, indicating that it is not necessary to upload files. Skip the upload step and directly delete the local redundant segments after merging the TabletMeta.
Note: In order to prevent inconsistency between different replicas of the local Rowset, which will result in the inability to de duplicate the cold data on the remote cluster, the cold data compression operation only processes the segments on the remote cluster, not the local segments.
BE in USE_ REMOTE_ DATA type
Since the remote TabletMeta may add or delete Rowsets, the local TabletMeta may not match the remote TabletMeta, that is, the specified file cannot be found when reading the remote data. In this case, you need to trigger the TabletMeta synchronization operation to merge the remote TabletMeta, and then continue the reading operation.
Scheduling
specific implementation steps and approximate scheduling.