[
https://issues.apache.org/jira/browse/HDDS-5359?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hanisha Koneru updated HDDS-5359:
---------------------------------
Description:
Here are default column family data of two different container replicas，
"#BCSID" -> 1354765
"#BLOCKCOUNT" -> -21
"#BYTESUSED" -> 0
"#PENDINGDELETEBLOCKCOUNT" -> 78
"#delTX" -> 1141106
"#BCSID" -> 1895040
"#BLOCKCOUNT" -> -5
"#BYTESUSED" -> 0
"#PENDINGDELETEBLOCKCOUNT" -> 106
"#delTX" -> 1146817
Update:
The BlockCount is incremented only when the Stream is closed and not when the
BlockID is added to the DB. If the OutputStream was not closed properly or if,
for any reason, the client starts writing to a new pipeline before the full
block is written, it could lead to a Block being present in the container but
the key_count (BlockCount) not being incremented for it. When a block is
deleted from a container, the blockCount is also decremented. But if the
blockCount is wrong to start with, it could lead to a negative value.
When a block is deleted, usedBytes is decrement in memory first after deleting
a chunk. And even if the chunkFile does not exist (already deleted), the
usedBytes is decremented. This could lead to usedBytes being decremented
multiple times for the same chunk and causing the total usedBytes metadata in
the DB to become negative. Once all the chunks in all the blocks in that
iteration of BlockDeletingService task are deleted, only then is the usedBytes
updated in containerDB by taking the in-memory value. This Jira proposes to
first update the DB with correct usedBytes (calculated from the BlockInfo after
all chunks are deleted) and then update the in-memory metadata. This is the
update sequence logic followed for all other state updates.
Also, when a chunk is overwritten, then it is assumed that the size of the
chunk remains the same. But it’s possible to overwrite more data into the chunk
than originally present. In this case, the used_bytes should be updated with
difference in the chunkSizes. (Adding this as a TODO).
was:
Here are default column family data of two different container replicas，
"#BCSID" -> 1354765
"#BLOCKCOUNT" -> -21
"#BYTESUSED" -> 0
"#PENDINGDELETEBLOCKCOUNT" -> 78
"#delTX" -> 1141106
"#BCSID" -> 1895040
"#BLOCKCOUNT" -> -5
"#BYTESUSED" -> 0
"#PENDINGDELETEBLOCKCOUNT" -> 106
"#delTX" -> 1146817
Update:
The BlockCount is incremented only when the Stream is closed and not when the
BlockID is added to the DB. If the OutputStream was not closed properly or if,
for any reason, the client starts writing to a new pipeline before the full
block is written, it could lead to a Block being present in the container but
the _key_count_ (BlockCount) not being incremented for it.
In the general case, the _used_bytes_ metadata is updated correctly i.e.
whenever a chunk is written and putBlock is called. But when a chunk is
overwritten, then it is assumed that the size of the chunk remains the same.
But it’s possible to overwrite more data into the chunk than originally
present. In this case, the _used_bytes_ should be updated with difference in
the chunkSizes.
When blocks are deleted from a container, the _key_count_ and _used_bytes_ is
decremented accordingly. But if these values were incorrect (less than the
actual), then deleting blocks could lead to negative values for these metadata.
> Incorrect BLOCKCOUNT and BYTESUSED in container DB
> --------------------------------------------------
>
> Key: HDDS-5359
> URL: https://issues.apache.org/jira/browse/HDDS-5359
> Project: Apache Ozone
> Issue Type: Bug
> Reporter: Sammi Chen
> Assignee: Hanisha Koneru
> Priority: Major
> Attachments: negative.txt
>
>
> Here are default column family data of two different container replicas，
> "#BCSID" -> 1354765
> "#BLOCKCOUNT" -> -21
> "#BYTESUSED" -> 0
> "#PENDINGDELETEBLOCKCOUNT" -> 78
> "#delTX" -> 1141106
> "#BCSID" -> 1895040
> "#BLOCKCOUNT" -> -5
> "#BYTESUSED" -> 0
> "#PENDINGDELETEBLOCKCOUNT" -> 106
> "#delTX" -> 1146817
>
> Update:
> The BlockCount is incremented only when the Stream is closed and not when the
> BlockID is added to the DB. If the OutputStream was not closed properly or
> if, for any reason, the client starts writing to a new pipeline before the
> full block is written, it could lead to a Block being present in the
> container but the key_count (BlockCount) not being incremented for it. When a
> block is deleted from a container, the blockCount is also decremented. But if
> the blockCount is wrong to start with, it could lead to a negative value.
> When a block is deleted, usedBytes is decrement in memory first after
> deleting a chunk. And even if the chunkFile does not exist (already deleted),
> the usedBytes is decremented. This could lead to usedBytes being decremented
> multiple times for the same chunk and causing the total usedBytes metadata in
> the DB to become negative. Once all the chunks in all the blocks in that
> iteration of BlockDeletingService task are deleted, only then is the
> usedBytes updated in containerDB by taking the in-memory value. This Jira
> proposes to first update the DB with correct usedBytes (calculated from the
> BlockInfo after all chunks are deleted) and then update the in-memory
> metadata. This is the update sequence logic followed for all other state
> updates.
> Also, when a chunk is overwritten, then it is assumed that the size of the
> chunk remains the same. But it’s possible to overwrite more data into the
> chunk than originally present. In this case, the used_bytes should be updated
> with difference in the chunkSizes. (Adding this as a TODO).
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
