[DNM] Wal debug 6.29.tikv #356
Conversation
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
|
@tonyxuqqi What is the WAL recovery mode in tikv ? I guess that it should be kTolerateCorruptedTailRecords. |
|
which format is tikv using. Seems like rocksdb wal has two formats : legacy and recycle. Recycle format has more debug information (log number) . https://github.com/facebook/rocksdb/wiki/Write-Ahead-Log-File-Format |
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
| "Start writing to WAL: [%" PRIu64 " ]", | ||
| log_writer->get_log_number()); | ||
| } | ||
| if (log_writer->get_log_number() != logs_.back().number) { |
There was a problem hiding this comment.
@tonyxuqqi : At the beginning of the write, writer will reference the last WAL only. The race condition happens, while the write is ongoing. We should check this condition after the write.
| @@ -1488,7 +1501,7 @@ IOStatus DBImpl::WriteToWAL(const WriteThread::WriteGroup& write_group, | |||
|
|
|||
| uint64_t log_size; | |||
| io_s = WriteToWAL(*merged_batch, log_writer, log_used, &log_size, | |||
| log_file_number_size); | |||
| log_file_number_size, 1); | |||
| if (to_be_cached_state) { | |||
There was a problem hiding this comment.
Can you also add a lot when the fsync happens in writeToWAL
rocksdb/db/db_impl/db_impl_write.cc
Line 1519 in 20225a4
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
This reverts commit 844300e.
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
Signed-off-by: Qi Xu <tonyxuqqi@outlook.com>
To debug WAL corruption issues