Our use of protobuf helps mitigate a lot of compatibility concerns, but there still can be situations that require careful coding on our part. When adding a new field to a protobuf message, the client might need to do a null check, even if the server-side implementation in the new version always populates the field. When adding a whole new RPC endpoint, the client might need to consider the possibility that the RPC endpoint isn't there on an old server, and degrade gracefully after the RPC fails. The original issue in MAPREDUCE-4052 concerned the script commands passed in a YARN container submission, where protobuf doesn't provide any validation beyond the fact that they're strings.
Forward compatibility is harder than backward compatibility, and testing is a big challenge. Our test suites in the Hadoop repo don't cover this. Does anyone know if anything in Bigtop tries to run with mixed versions? I agree that we need to make it clear in the language that upgrading client alone is insufficient to get access to new server-side features, including new YARN APIs. Thanks for the suggestions, Steve. Chris Nauroth Hortonworks http://hortonworks.com/ On Thu, Mar 20, 2014 at 5:53 AM, Steve Loughran <ste...@hortonworks.com>wrote: > I'm clearly supportive of this, though of course the testing costs needed > to back up the assertion make it more expensive than just a statement. > > Two issues > > -we'd need to make clear that new cluster features that a client can invoke > won't be available. You can't expect snapshot or symlink support running > against a -2.2.0 cluster, even if the client supports it. > > -in YARN, there are no guarantees that an app compiled against later YARN > APIs will work in old clusters. Because YARN apps upload themselves to the > server, and run with their hadoop, hdfs & yarn libraries. We have to do a > bit of introspection in our code already to support this situation. The > compatibility doc would need to be clear on that too: "YARN apps that use > new APIs (including new fields in datastructures) can expect link > exceptions" > > > > > > On 20 March 2014 04:25, Vinayakumar B <vinayakuma...@huawei.com> wrote: > > > +1, I agree with your point Chris. It depends on the client application > > how they using the hdfs jars in their classpath. > > > > As implementation already supports the compatibility (through protobuf), > > No extra code changes required to support new Client + old server. > > > > I feel it will be good to explicitly mention about the compatibility of > > existing APIs in both versions. > > > > Anyway this is not applicable for the new APIs in latest client and this > > is understood. We can make it explicit in the document though. > > > > > > Regards, > > Vinayakumar B > > > > -----Original Message----- > > From: Chris Nauroth [mailto:cnaur...@hortonworks.com] > > Sent: 20 March 2014 05:36 > > To: common-...@hadoop.apache.org > > Cc: mapreduce-...@hadoop.apache.org; hdfs-...@hadoop.apache.org; > > yarn-dev@hadoop.apache.org > > Subject: Re: [DISCUSS] Clarification on Compatibility Policy: Upgraded > > Client + Old Server > > > > I think this kind of compatibility issue still could surface for HDFS, > > particularly for custom applications (i.e. something not executed via > > "hadoop jar" on a cluster node, where the client classes ought to be > > injected into the classpath automatically). Running DistCP between 2 > > clusters of different versions could result in a 2.4.0 client calling a > > 2.3.0 NameNode. Someone could potentially pick up the 2.4.0 WebHDFS > > client as a dependency and try to use it to make HTTP calls to a 2.3.0 > HDFS > > cluster. > > > > Chris Nauroth > > Hortonworks > > http://hortonworks.com/ > > > > > > > > On Wed, Mar 19, 2014 at 4:28 PM, Vinod Kumar Vavilapalli < > > vino...@apache.org > > > wrote: > > > > > It makes sense only for YARN today where we separated out the clients. > > > HDFS is still a monolithic jar so this compatibility issue is kind of > > > invalid there. > > > > > > +vinod > > > > > > On Mar 19, 2014, at 1:59 PM, Chris Nauroth <cnaur...@hortonworks.com> > > > wrote: > > > > > > > I'd like to discuss clarification of part of our compatibility > policy. > > > > Here is a link to the compatibility documentation for release 2.3.0: > > > > > > > > > > > http://hadoop.apache.org/docs/r2.3.0/hadoop-project-dist/hadoop-common > > > /Compatibility.html#Wire_compatibility > > > > > > > > For convenience, here are the specific lines in question: > > > > > > > > Client-Server compatibility is required to allow users to continue > > > > using the old clients even after upgrading the server (cluster) to a > > > > later version (or vice versa). For example, a Hadoop 2.1.0 client > > > > talking to a Hadoop 2.3.0 cluster. > > > > > > > > Client-Server compatibility is also required to allow upgrading > > > individual > > > > components without upgrading others. For example, upgrade HDFS from > > > version > > > > 2.1.0 to 2.2.0 without upgrading MapReduce. > > > > > > > > Server-Server compatibility is required to allow mixed versions > > > > within an active cluster so the cluster may be upgraded without > > > > downtime in a > > > rolling > > > > fashion. > > > > > > > > Notice that there is no specific mention of upgrading the client > > > > ahead of the server. (There is no clause for "upgraded client + old > > > > server".) Based on my experience, this is a valid use case when a > > > > user wants to > > > pick > > > > up a client-side bug fix ahead of the cluster administrator's > > > > upgrade schedule. > > > > > > > > Is it our policy to maintain client compatibility with old clusters > > > within > > > > the same major release? I think many of us have assumed that the > > > > answer > > > is > > > > yes and coded our new features accordingly, but it isn't made > > > > explicit in the documentation. Do we all agree that the answer is > > > > yes, or is it possibly up for debate depending on the change in > > > > question? In RFC 2119 lingo, is it a MUST or a SHOULD? Either way, > > > > I'd like to update the > > > policy > > > > text to make our decision clear. After we have consensus, I can > > > volunteer > > > > to file an issue and patch the text of the policy. > > > > > > > > This discussion started initially in MAPREDUCE-4052, which involved > > > > changing our scripting syntax for MapReduce YARN container > submissions. > > > We > > > > settled the question there by gating the syntax change behind a > > > > configuration option. By default, it will continue using the > > > > existing syntax currently understood by the pre-2.4.0 NodeManager, > > > > thus preserving compatibility. We wanted to open the policy > > > > question for wider > > > discussion > > > > though. > > > > > > > > Thanks, everyone. > > > > > > > > Chris Nauroth > > > > Hortonworks > > > > http://hortonworks.com/ > > > > > > > > -- > > > > CONFIDENTIALITY NOTICE > > > > NOTICE: This message is intended for the use of the individual or > > > > entity > > > to > > > > which it is addressed and may contain information that is > > > > confidential, privileged and exempt from disclosure under applicable > > > > law. If the reader of this message is not the intended recipient, > > > > you are hereby notified > > > that > > > > any printing, copying, dissemination, distribution, disclosure or > > > > forwarding of this communication is strictly prohibited. If you have > > > > received this communication in error, please contact the sender > > > immediately > > > > and delete it from your system. Thank You. > > > > > > > > > -- > > > CONFIDENTIALITY NOTICE > > > NOTICE: This message is intended for the use of the individual or > > > entity to which it is addressed and may contain information that is > > > confidential, privileged and exempt from disclosure under applicable > > > law. If the reader of this message is not the intended recipient, you > > > are hereby notified that any printing, copying, dissemination, > > > distribution, disclosure or forwarding of this communication is > > > strictly prohibited. If you have received this communication in error, > > > please contact the sender immediately and delete it from your system. > > Thank You. > > > > > > > -- > > CONFIDENTIALITY NOTICE > > NOTICE: This message is intended for the use of the individual or entity > > to which it is addressed and may contain information that is > confidential, > > privileged and exempt from disclosure under applicable law. If the reader > > of this message is not the intended recipient, you are hereby notified > that > > any printing, copying, dissemination, distribution, disclosure or > > forwarding of this communication is strictly prohibited. If you have > > received this communication in error, please contact the sender > immediately > > and delete it from your system. Thank You. > > > > -- > CONFIDENTIALITY NOTICE > NOTICE: This message is intended for the use of the individual or entity to > which it is addressed and may contain information that is confidential, > privileged and exempt from disclosure under applicable law. If the reader > of this message is not the intended recipient, you are hereby notified that > any printing, copying, dissemination, distribution, disclosure or > forwarding of this communication is strictly prohibited. If you have > received this communication in error, please contact the sender immediately > and delete it from your system. Thank You. > -- CONFIDENTIALITY NOTICE NOTICE: This message is intended for the use of the individual or entity to which it is addressed and may contain information that is confidential, privileged and exempt from disclosure under applicable law. If the reader of this message is not the intended recipient, you are hereby notified that any printing, copying, dissemination, distribution, disclosure or forwarding of this communication is strictly prohibited. If you have received this communication in error, please contact the sender immediately and delete it from your system. Thank You.
