Hi Haohui, I'm glad to know GRPC and it sounds cool. I think it's a good proposal to suggest Hadoop IPC/RPC upgrading to GRPC.
We haven't evaluated GRPC for the question of RPC encryption optimization because it's another story. It's not an overlap for the optimization work because even if we use GRPC, the RPC protocol messages still need to go through the stack of SASL/GSSAPI/Kerberos. What's desired here is not to re-implement any RPC layer, or the stack, but is to optimize the stack, by possibly implementing and plugin-ing new SASL or GSSAPI mechanism. Hope this clarifying helps. Thanks. Regards, Kai -----Original Message----- From: Haohui Mai [mailto:ricet...@gmail.com] Sent: Sunday, February 28, 2016 3:02 AM To: common-dev@hadoop.apache.org Subject: Re: Introduce Apache Kerby to Hadoop Have we evaluated GRPC? A robust RPC requires significant effort. Migrating to GRPC can save ourselves a lot of headache. Haohui On Sat, Feb 27, 2016 at 1:35 AM Andrew Purtell <andrew.purt...@gmail.com> wrote: > I get a excited thinking about the prospect of better performance with > auth-conf QoP. HBase RPC is an increasingly distant fork but still > close enough to Hadoop in that respect. Our bulk data transfer > protocol isn't a separate thing like in HDFS, which avoids a SASL > wrapped implementation, so we really suffer when auth-conf is > negotiated. You'll see the same impact where there might be a high > frequency of NameNode RPC calls or similar still. Throughput drops 3-4x, or > worse. > > > On Feb 22, 2016, at 4:56 PM, Zheng, Kai <kai.zh...@intel.com> wrote: > > > > Thanks for the confirm and further inputs, Steve. > > > >>> the latter would dramatically reduce the cost of wire-encrypting IPC. > > Yes to optimize Hadoop IPC/RPC encryption is another opportunity > > Kerby > can help with, it's possible because we may hook Chimera or AES-NI > thing into the Kerberos layer by leveraging the Kerberos library. As > it may be noted, HADOOP-12725 is on the going for this aspect. There > may be good result and further update on this recently. > > > >>> For now, I'd like to see basic steps -upgrading minkdc to krypto, > >>> see > how it works. > > Yes, starting with this initial steps upgrading MiniKDC to use Kerby > > is > the right thing we could do. After some interactions with Kerby > project, we may have more ideas how to proceed on the followings. > > > >>> Long term, I'd like Hadoop 3 to be Kerby-ized > > This sounds great! With necessary support from the community like > feedback and patch reviewing, we can speed up the related work. > > > > Regards, > > Kai > > > > -----Original Message----- > > From: Steve Loughran [mailto:ste...@hortonworks.com] > > Sent: Monday, February 22, 2016 6:51 PM > > To: common-dev@hadoop.apache.org > > Subject: Re: Introduce Apache Kerby to Hadoop > > > > > > > > I've discussed this offline with Kai, as part of the "let's fix > kerberos" project. Not only is it a better Kerberos engine, we can do > more diagnostics, get better algorithms and ultimately get better APIs > for doing Kerberos and SASL —the latter would dramatically reduce the > cost of wire-encrypting IPC. > > > > For now, I'd like to see basic steps -upgrading minkdc to krypto, > > see > how it works. > > > > Long term, I'd like Hadoop 3 to be Kerby-ized > > > > > >> On 22 Feb 2016, at 06:41, Zheng, Kai <kai.zh...@intel.com> wrote: > >> > >> Hi folks, > >> > >> I'd like to mention Apache Kerby [1] here to the community and > >> propose > to introduce the project to Hadoop, a sub project of Apache Directory > project. > >> > >> Apache Kerby is a Kerberos centric project and aims to provide a > >> first > Java Kerberos library that contains both client and server supports. > The relevant features include: > >> It supports full Kerberos encryption types aligned with both MIT > >> KDC and MS AD; Client APIs to allow to login via password, > >> credential cache, keytab file and etc.; Utilities for generate, > >> operate and inspect keytab and credential cache files; A simple KDC > >> server that borrows some ideas from Hadoop-MiniKDC and can be used > >> in tests but with minimal overhead in external dependencies; A > >> brand new token > mechanism is provided, can be experimentally used, using it a JWT > token can be used to exchange a TGT or service ticket; Anonymous > PKINIT support, can be experientially used, as the first Java library > that supports the Kerberos major extension. > >> > >> The project stands alone and is ensured to only depend on JRE for > easier usage. It has made the first release (1.0.0-RC1) and 2nd > release > (RC2) is upcoming. > >> > >> > >> As an initial step, this proposal suggests using Apache Kerby to > upgrade the existing codes related to ApacheDS for the Kerberos support. > The advantageous: > >> > >> 1. The kerby-kerb library is all the need, which is purely in Java, > >> SLF4J is the only dependency, the whole is rather small; > >> > >> 2. There is a SimpleKDC in the library for test usage, which > >> borrowed the MiniKDC idea and implemented all the support existing in > >> MiniKDC. > >> We had a POC that rewrote MiniKDC using Kerby SimpleKDC and it > >> works fine; > >> > >> 3. Full Kerberos encryption types (many of them are not available > >> in JRE but supported by major Kerberos vendors) and more > >> functionalities like credential cache support; > >> > >> 4. Perhaps the most concerned, Hadoop MiniKDC and etc. depend on > >> the old Kerberos implementation in Directory Server project, but > >> the implementation is stopped being maintained. Directory project > >> has a plan to replace the implementation using Kerby. MiniKDC can > >> use Kerby directly to simplify the deps; > >> > >> 5. Extensively tested with all kinds of unit tests, already being > >> used for some time (like PSU), even in production environment; > >> > >> 6. Actively developed, and can be fixed and released in time if > necessary, separately and independently from other components in > Apache Directory project. By actively developing Apache Kerby and now > applying it to Hadoop, our side wish to make the Kerberos deploying, > troubleshooting and further enhancement can be much easier and thereafter > possible. > >> > >> > >> > >> Wish this is a good beginning, and eventually Apache Kerby can > >> benefit > other projects in the ecosystem as well. > >> > >> > >> > >> This Kerberos related work is actually a long time effort led by > >> Weihua > Jiang in Intel, and had been kindly encouraged by Andrew Purtell, > Steve Loughran, Gangumalla Uma, Andrew Wang and etc., thanks a lot for > their great discussions and inputs in the past. > >> > >> > >> > >> Your feedback is very welcome. Thanks in advance. > >> > >> > >> > >> [1] https://github.com/apache/directory-kerby > >> > >> > >> > >> Regards, > >> > >> Kai > > >
