Xiaolin Ha created HBASE-26552: ---------------------------------- Summary: Introduce retry to logroller when encounters IOException Key: HBASE-26552 URL: https://issues.apache.org/jira/browse/HBASE-26552 Project: HBase Issue Type: Improvement Components: wal Affects Versions: 2.0.0, 3.0.0-alpha-1 Reporter: Xiaolin Ha Assignee: Xiaolin Ha
When calling RollController#rollWal in AbstractWALRoller, the regionserver may abort when encounters exception, {code:java} ... } catch (FailedLogCloseException | ConnectException e) { abort("Failed log close in log roller", e); } catch (IOException ex) { // Abort if we get here. We probably won't recover an IOE. HBASE-1132 abort("IOE in log roller", ex instanceof RemoteException ? ((RemoteException) ex).unwrapRemoteException() : ex); } catch (Exception ex) { LOG.error("Log rolling failed", ex); abort("Log rolling failed", ex); } {code} I think we should support retry of rollWal here to avoid recovering the service by killing regionserver. The restart of regionserver is costly and very not friendly to the availability. I find that when creating new writer for the WAL in FanOutOneBlockAsyncDFSOutputHelper#createOutput, it supports retry to addBlock by setting this config "hbase.fs.async.create.retries". But the initialization of new WAL writer also includes flushing the write buffer flush and waiting until it is completed by AsyncProtobufLogWriter#writeMagicAndWALHeader, which can also fail by some hardware reasons. The regionserver connected to the datanodes after addBlock, but that not means the magic and header can be flushed successfully. {code:java} protected long writeMagicAndWALHeader(byte[] magic, WALHeader header) throws IOException { return write(future -> { output.write(magic); try { header.writeDelimitedTo(asyncOutputWrapper); } catch (IOException e) { // should not happen throw new AssertionError(e); } addListener(output.flush(false), (len, error) -> { if (error != null) { future.completeExceptionally(error); } else { future.complete(len); } }); }); }{code} We have found that in our production clusters, there exists aborting of regionservers that caused by "IOE in log roller". And the practice in our clusters is that just one more retry of rollWal can make the WAL roll complete and continue serving. -- This message was sent by Atlassian Jira (v8.20.1#820001)