Support for Phoenix

[jtaylor-sfdc]

Great job on the transaction support! It would be interesting to have a version of Haeinsa that works with Phoenix.

[satoshi75nakamoto]

👍

[satoshi75nakamoto]

I'd love to see Salesforce Phoenix support Haeinsa.

[rrmckinley]

👍

[jmlvanre]

This support would be very useful.

[eincs]

Hi everyone! Sorry for late reply. Thanks for good suggestion.
I would be very excited and pleased, if Haeinsa helps Phoenix to support transaction.
I'm now studying Phoenix for few days and I'll deeply see how Phoenix implemented and finding good way to Phoenix and Haeinsa works together. If anyone who have the good suggestion about this issues (like implementation ideas), Please let me know.
Thanks!

[satoshi75nakamoto]

Thanks @eincs.

[jtaylor-sfdc]

Thanks, @eincs. I'm happy to answer any questions you have about Phoenix.

[satoshi75nakamoto]

@eincs What do you think of Phoenix so far?

[eincs]

@preillyme I'll study Phoenix this weekend.
Do you have any opinion, such as how to integrate Pheonix and Haeinsa?

[eincs]

@jtaylor-sfdc @preillyme Here is my thought on this issue.

Transaction algorithm of Haeinsa relies on checkAndPut and checkAndDelete of HBase client. Those operations are carefully managed by several classes in Haeinsa, such as HaeinsaTable. Note that Haeinsa do not uses any coprocessor and never modifies Region Server.

In my opinion, to use Haeinsa in Pheonix for transaction, Phoenix and Haeinsa should both modified. Since Haeinsa relies on client-side algorithm, Pheonix should use Haeinsa client internally. This means codes of Pheonix should be modified quite a lot. And also there is some functions which Haeinsa doesn't support (which are supported on Hbase client) These should be fixed in Haeinsa-side.

For example, I see Pheonix use batch(Rows) operation on MutationState class. And Pheonix creates Scan operation in ScanUtils. Following can be problem in Haeinsa:

• Haeinsa don't support batch(Rows)
• Haeinsa operations, such as HaeinsaPut, HaeinsaDelete, do not have setAttribute method: since pheonix uses attribute for coprocessors, this is neccessary.
• Haeinsa don't support filter in HaeinsaScan.

I'm totally agree with setAttribite should be implemented on Haeinsa. And I also think it is reasonable that implementing filter in HaeinsaScan operation. But I don't sure about implementing batch(Rows) because Haeinsa executes all operation in the transaction batch in commit phase. We can discuss about other problems.

And note that Haeinsa is designed to have good performance in small number of row in single transaction. Haeinsa works best on scheme which is similar to facebook messaging system, which all the data for user are stored in single row. You can see blog post1 from facebook engineering team. This means performance of Haeinsa can be not so good according to schema of the table.

P.S.
I studied Pheonix last weekend, but there can be some misunderstand, so please let me know if there is error.

[eincs]

I added related issues. See #10 and #11.

[satoshi75nakamoto]

@eincs thanks for adding those other issues that will aid in Phoenix support.