Good proposal.
It is important that this is disabled by default otherwise we cannot easily
support the rollback

Apart from that I don't have other comments



Enrico

Il Sab 5 Mar 2022, 11:22 PengHui Li <peng...@apache.org> ha scritto:

> Hi Zixuan,
>
> We should add the compatibility part to the proposal.
> And should also provide steps to roll back to the old version which
> enabled the compression in the new version.
>
> I don't have objections to the proposal, and we have done
> the same enhancement for topic metadata
> https://github.com/apache/pulsar/pull/11490,
> and this proposal also follows the same way.
>
> Thanks,
> Penghui
>
> On Thu, Mar 3, 2022 at 10:26 AM Zixuan Liu <node...@gmail.com> wrote:
>
> > Thank you for your feedback.
> >
> > Forward compatibility is required:
> >
> > 1. Get CursorInfo from the metadata store: We will check the cursor data
> > header, if it is compressed, we will parse the bytes data by compressed
> > format, otherwise we will parse the cursor data directly by the original
> > way.
> >
> > 2. Add/Update CursorInfo to the metadata store: The default is to use
> > compression if the compression type is specified, otherwise we will put
> > this data to the metadata store directly.
> >
> >
> >
> > > 2022年3月2日 下午11:48,Zike Yang <zky...@streamnative.io.INVALID> 写道:
> > >
> > > Hi, Zixuan
> > > Thanks for creating this PIP.  Here are my thoughts.
> > >
> > >> CursorInfo compression format
> > >>
> > >> [MAGIC_NUMBER] + [METADATA_SIZE] + [METADATA_PAYLOAD] +
> > [MANAGED_CURSOR_INFO_PAYLOAD]
> > >>
> > >> MAGIC_NUMBER: Ox4779
> > >
> > > Since we change the ManagedCursorInfo data format here. How do we
> > > handle the old data format in the ZK. Could you explain the
> > > compatibility for this PIP?
> > >
> > > Thanks,
> > > Zike Yang
> > >
> > > On Wed, Mar 2, 2022 at 3:34 PM Zixuan Liu <node...@gmail.com <mailto:
> > node...@gmail.com>> wrote:
> > >>
> > >> Hi Pulsar Community,
> > >>
> > >> I create a proposal that ManagedCursorInfo compression. The proposal
> > can be found: https://github.com/apache/pulsar/issues/14529 <
> > https://github.com/apache/pulsar/issues/14529> <
> > https://github.com/apache/pulsar/issues/14529 <
> > https://github.com/apache/pulsar/issues/14529>>
> > >>
> > >> Thanks,
> > >> Zixuan
> > >>
> > >> ------------------
> > >>
> > >> Motivation
> > >>
> > >> The cursor data is managed by ZooKeeper/etcd metadata store. When
> > cursor data becomes more and more, the data size will increase and will
> > take a lot of time to pull the data. Therefore, it is necessary to add
> > compression for the cursor, which can reduce the size of data and reduce
> > the time of pulling data.
> > >>
> > >> Goal
> > >>
> > >> Support use the LZ4/ZLIB/ZSTD/SNAPPY to compress the
> ManagedCursorInfo.
> > >>
> > >> Implementation
> > >>
> > >> CursorInfo compression format
> > >>
> > >> [MAGIC_NUMBER] + [METADATA_SIZE] + [METADATA_PAYLOAD] +
> > [MANAGED_CURSOR_INFO_PAYLOAD]
> > >>
> > >> MAGIC_NUMBER: Ox4779
> > >>
> > >> METADATA
> > >> Add a named ManagedCursorInfoMetadata message to MLDataFormats.proto
> > >>
> > >> message ManagedCursorInfoMetadata {
> > >>    required CompressionType compressionType = 1;
> > >>    required int32 uncompressedSize = 2;
> > >> }
> > >> CursorInfo compression and decompression design
> > >>
> > >> Currently, these compressions types have been defined and implemented
> > by Pulsar, we only need to deal with compression and decompression of the
> > ManagedCursorInfo data:
> > >>
> > >> Get CursorInfo from the metadata store
> > >> We will check the cursor data header, if it is compressed, we will
> > parse the bytes data by compressed format, otherwise by the original way.
> > >>
> > >> Add/Update CursorInfo to the metadata store
> > >> The default is to use compression if the compression type is
> specified.
> > >>
> > >> CursorInfo compression type configuration
> > >>
> > >> Add managedCursorInfoCompressionType in
> > org.apache.pulsar.broker.ServiceConfiguration and
> > org.apache.bookkeeper.mledger.ManagedLedgerFactoryConfig.
> > >>
> > >>
> > >
> > >
> > > --
> > > Zike Yang
> >
> >
>

Reply via email to