Yeah they are very tricky and have to be integrated with Spark's checkpoint
mechanism as well - I guess that's why this mail thread had been quiet
after some time.

Along with these questions, there might be also some edge-cases which we
have to deal with 2PC approach: suppose a batch got into commit phase and
some tasks succeeded to commit, and strangely a task becomes unable to
commit (transaction timeout, crashed and lost handle on transaction, etc.).
In this case Spark would be forced to rerun failed batch, falling back to
at-least-once.

2PC via "WAL & transaction per task" might do the trick for answering these
questions and the edge-case, but yes that's writing twice (even required to
write it to durable storage regardless of the kind of sink), though final
write would be still scalable.

On Fri, Dec 6, 2019 at 1:58 PM Wenchen Fan <cloud0...@gmail.com> wrote:

> I also share the concerns of "writing twice", which hurts performance a
> lot. What's worse, the final write may not be scalable, like writing the
> staging table to the final table.
>
> If the sink itself doesn't support global transaction, but only local
> transaction (e.g. kafla), using staging tables seems the only way to make
> the write atomic. But if we accept "eventually consistent", we can use 2PC
> to avoid "writing twice". i.e. wait for all the tasks to finish
> writing data, then ask them to commit at the same time.
>
> There are 2 open questions we need to answer:
> 1. How to make sure all tasks are launched at the same time to implement
> 2PC? barrier execution?
> 2. To reach "eventually consistent", we must retry the job until successe.
> How shall we guarantee the job retry?
>
> On Fri, Oct 19, 2018 at 12:25 PM Jungtaek Lim <kabh...@gmail.com> wrote:
>
>> Sorry to resurrect this old and long thread: we have been struggling with
>> Kafka end-to-end exactly-once support, and couldn't find any approach which
>> can get both things, transactional and scalable.
>>
>> If we tolerate scalability, we can let writers to write to staging topic
>> within individual transaction, and once all writers are succeed to write,
>> let coordinator re-read topic and write to final topic in its transaction.
>> (Coordinator should be able to get offsets to read and skip offsets which
>> are being left due to failed trial of batch, so shouldn't be read-committed
>> while reading.)
>>
>> This will result in twice of data writing as well as all rows being
>> published within single thread in final step, which most of the cases we
>> can't tolerate. Moreover, repartitioning to 1 and write with enabling
>> transaction might do the same thing better: staging topic vs shuffle.
>>
>> If we tolerate transaction (I meant "global transaction") and allow
>> "eventually consistent" - allow part(s) of output being seen to client side
>> at specific point (2PC has same limitation), there could be some approaches
>> which might be tricky but work.
>>
>> Even external storages support transaction, normally it doesn't mean they
>> are supporting global transaction or grouped transactions. The boundary of
>> transaction is mostly limited to the one client (connection), and once we
>> want to write from multiple tasks, we will encounter same issue on these
>> external storages, except the cases (like HDFS) which can "move" its data
>> from staging to final destination within storage.
>>
>> So could we consider lessen the contract on DataSource V2 writer, or have
>> a new representation of guarantee for such case so it is not "fully
>> transactional" but another kind of "exactly-once" and not "at-least-once"?
>>
>> Thanks,
>> Jungtaek Lim (HeartSaVioR)
>>
>> 2018년 9월 14일 (금) 오전 12:08, Thakrar, Jayesh <jthak...@conversantmedia.com>님이
>> 작성:
>>
>>> Agree on the “constraints” when working with Cassandra.
>>>
>>> But remember, this is a weak attempt to make two non-transactional
>>> systems appear to the outside world as a transactional system.
>>>
>>> Scaffolding/plumbing/abstractions will have to be created in the form of
>>> say, a custom data access layer.
>>>
>>>
>>>
>>> Anyway, Ross is trying to get some practices used by other adopters of
>>> the V2 API while trying to implement a driver/connector for MongoDB.
>>>
>>>
>>>
>>> Probably views can be used similar to partitions in mongoDB?
>>>
>>> Essentially each batch load goes into a separate mongoDB table and will
>>> result in view redefinition after a successful load.
>>>
>>> And finally to avoid too many tables in a view, you may have to come up
>>> with a separate process to merge the underlying tables on a periodic basis.
>>>
>>> It gets messy and probably moves you towards a write-once only tables,
>>> etc.
>>>
>>>
>>>
>>> Finally using views in a generic mongoDB connector may not be good and
>>> flexible enough.
>>>
>>>
>>>
>>>
>>>
>>> *From: *Russell Spitzer <russell.spit...@gmail.com>
>>> *Date: *Tuesday, September 11, 2018 at 9:58 AM
>>> *To: *"Thakrar, Jayesh" <jthak...@conversantmedia.com>
>>> *Cc: *Arun Mahadevan <ar...@apache.org>, Jungtaek Lim <kabh...@gmail.com>,
>>> Wenchen Fan <cloud0...@gmail.com>, Reynold Xin <r...@databricks.com>,
>>> Ross Lawley <ross.law...@gmail.com>, Ryan Blue <rb...@netflix.com>, dev
>>> <dev@spark.apache.org>, "dbis...@us.ibm.com" <dbis...@us.ibm.com>
>>>
>>>
>>> *Subject: *Re: DataSourceWriter V2 Api questions
>>>
>>>
>>>
>>> That only works assuming that Spark is the only client of the table. It
>>> will be impossible to force an outside user to respect the special metadata
>>> table when reading so they will still see all of the data in transit.
>>> Additionally this would force the incoming data to only be written into new
>>> partitions which is not simple to do from a C* perspective as balancing the
>>> distribution of new rows would be non trivial. If we had to do something
>>> like this we would basically be forced to write to some disk format first
>>> and then when we move the data into C* we still have the same problem that
>>> we started with.
>>>
>>>
>>>
>>> On Tue, Sep 11, 2018 at 9:41 AM Thakrar, Jayesh <
>>> jthak...@conversantmedia.com> wrote:
>>>
>>> So if Spark and the destination datastore are both non-transactional,
>>> you will have to resort to an external mechanism for “transactionality”.
>>>
>>>
>>>
>>> Here are some options for both RDBMS and non-transaction datastore
>>> destination.
>>>
>>> For now assuming that Spark is used in batch mode (and not streaming
>>> mode).
>>>
>>>
>>>
>>> *RDBMS Options*
>>>
>>> Use staging table as discussed in the thread.
>>>
>>>
>>>
>>> As an extension of the above, use partitioned destination tables and
>>> load data into a staging table and then use partition management to include
>>> the staging table into the partitioned table.
>>>
>>> This this implies a partition per Spark batch run.
>>>
>>>
>>>
>>> *Non-transactional Datastore Options*
>>>
>>> Use another metadata table.
>>>
>>> Load the data into a staging table equivalent or even Cassandra
>>> partition(s).
>>>
>>> Start the transaction by making a “start of transaction” entry into the
>>> metadata table along with partition keys to be populated.
>>> As part of Spark batch commit, update the metadata entry with
>>> appropriate details – e.g. partition load time, etc.
>>> In the event of a failed / incomplete batch, the metadata table entry
>>> will be incomplete and the corresponding partition keys can be dropped.
>>>
>>> So essentially you use the metadata table to load/drop/skip the data to
>>> be moved/retained into the final destination.
>>>
>>>
>>>
>>> *Misc*
>>>
>>> Another option is to use Spark to stage data into a filesystem
>>> (distributed, HDFS) and then use RDBMS utilities to transactionally load
>>> data into the destination table.
>>>
>>>
>>>
>>>
>>>
>>> *From: *Russell Spitzer <russell.spit...@gmail.com>
>>> *Date: *Tuesday, September 11, 2018 at 9:08 AM
>>> *To: *Arun Mahadevan <ar...@apache.org>
>>> *Cc: *Jungtaek Lim <kabh...@gmail.com>, Wenchen Fan <cloud0...@gmail.com>,
>>> Reynold Xin <r...@databricks.com>, Ross Lawley <ross.law...@gmail.com>,
>>> Ryan Blue <rb...@netflix.com>, dev <dev@spark.apache.org>, <
>>> dbis...@us.ibm.com>
>>>
>>>
>>> *Subject: *Re: DataSourceWriter V2 Api questions
>>>
>>>
>>>
>>> I'm still not sure how the staging table helps for databases which do
>>> not have such atomicity guarantees. For example in Cassandra if you wrote
>>> all of the data temporarily to a staging table, we would still have the
>>> same problem in moving the data from the staging table into the real table.
>>> We would likely have as similar a chance of failing and we still have no
>>> way of making the entire staging set simultaneously visible.
>>>
>>>
>>>
>>> On Tue, Sep 11, 2018 at 8:39 AM Arun Mahadevan <ar...@apache.org> wrote:
>>>
>>> >Some being said it is exactly-once when the output is eventually
>>> exactly-once, whereas others being said there should be no side effect,
>>> like consumer shouldn't see partial write. I guess 2PC is former, since
>>> some partitions can commit earlier while other partitions fail to commit
>>> for some time.
>>>
>>> Yes its more about guaranteeing atomicity like all partitions eventually
>>> commit or none commits. The visibility of the data for the readers is
>>> orthogonal (e.g setting the isolation levels like serializable for XA) and
>>> in general its difficult to guarantee that data across partitions are
>>> visible at once. The approach like staging table and global commit works in
>>> a centralized set up but can be difficult to do in a distributed manner
>>> across partitions (e.g each partition output goes to a different database)
>>>
>>>
>>>
>>> On Mon, 10 Sep 2018 at 21:23, Jungtaek Lim <kabh...@gmail.com> wrote:
>>>
>>> IMHO that's up to how we would like to be strict about "exactly-once".
>>>
>>>
>>>
>>> Some being said it is exactly-once when the output is eventually
>>> exactly-once, whereas others being said there should be no side effect,
>>> like consumer shouldn't see partial write. I guess 2PC is former, since
>>> some partitions can commit earlier while other partitions fail to commit
>>> for some time.
>>>
>>>
>>>
>>> Being said, there may be couple of alternatives other than the contract
>>> Spark provides/requires, and I'd like to see how Spark community wants to
>>> deal with others. Would we want to disallow alternatives, like "replay +
>>> deduplicate write (per a batch/partition)" which ensures "eventually"
>>> exactly-once but cannot ensure the contract?
>>>
>>>
>>>
>>> Btw, unless achieving exactly-once is light enough for given sink, I
>>> think the sink should provide both at-least-once (also optimized for the
>>> semantic) vs exactly-once, and let end users pick one.
>>>
>>>
>>>
>>> 2018년 9월 11일 (화) 오후 12:57, Russell Spitzer <russell.spit...@gmail.com>님이
>>> 작성:
>>>
>>> Why is atomic operations a requirement? I feel like doubling the amount
>>> of writes (with staging tables) is probably a tradeoff that the end user
>>> should make.
>>>
>>> On Mon, Sep 10, 2018, 10:43 PM Wenchen Fan <cloud0...@gmail.com> wrote:
>>>
>>> Regardless the API, to use Spark to write data atomically, it requires
>>>
>>> 1. Write data distributedly, with a central coordinator at Spark driver.
>>>
>>> 2. The distributed writers are not guaranteed to run together at the
>>> same time. (This can be relaxed if we can extend the barrier scheduling
>>> feature)
>>>
>>> 3. The new data is visible if and only if all distributed writers
>>> success.
>>>
>>>
>>>
>>> According to these requirements, I think using a staging table is the
>>> most common way and maybe the only way. I'm not sure how 2PC can help, we
>>> don't want users to read partial data, so we need a final step to commit
>>> all the data together.
>>>
>>>
>>>
>>> For RDBMS data sources, I think a simple solution is to ask users to
>>> coalesce the input RDD/DataFrame into one partition, then we don't need to
>>> care about multi-client transaction. Or using a staging table like Ryan
>>> described before.
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Tue, Sep 11, 2018 at 5:10 AM Jungtaek Lim <kabh...@gmail.com> wrote:
>>>
>>> > And regarding the issue that Jungtaek brought up, 2PC doesn't require
>>> tasks to be running at the same time, we need a mechanism to take down
>>> tasks after they have prepared and bring up the tasks during the commit
>>> phase.
>>>
>>>
>>>
>>> I guess we already got into too much details here, but if it is based on
>>> client transaction Spark must assign "commit" tasks to the executor which
>>> task was finished "prepare", and if it loses executor it is not feasible to
>>> force committing. Staging should come into play for that.
>>>
>>>
>>>
>>> We should also have mechanism for "recovery": Spark needs to ensure it
>>> finalizes "commit" even in case of failures before starting a new batch.
>>>
>>>
>>>
>>> So not an easy thing to integrate correctly.
>>>
>>> 2018년 9월 11일 (화) 오전 6:00, Arun Mahadevan <ar...@apache.org>님이 작성:
>>>
>>> >Well almost all relational databases you can move data in a
>>> transactional way. That’s what transactions are for.
>>>
>>>
>>>
>>> It would work, but I suspect in most cases it would involve moving data
>>> from temporary tables to the final tables
>>>
>>>
>>>
>>> Right now theres no mechanisms to let the individual tasks commit in a
>>> two-phase manner (Not sure if the CommitCordinator might help). If such an
>>> API is provided, the sources could use it as they wish (e.g. use XA support
>>> provided by mysql to implement it in a more efficient way than the driver
>>> moving from temp tables to destination tables).
>>>
>>>
>>>
>>> Definitely there are complexities involved, but I am not sure if the
>>> network partitioning comes into play here since the driver can act as the
>>> co-ordinator and can run in HA mode. And regarding the issue that Jungtaek
>>> brought up, 2PC doesn't require tasks to be running at the same time, we
>>> need a mechanism to take down tasks after they have prepared and bring up
>>> the tasks during the commit phase.
>>>
>>>
>>>
>>> Most of the sources would not need any of the above and just need a way
>>> to support Idempotent writes and like Ryan suggested we can enable this (if
>>> there are gaps in the current APIs).
>>>
>>>
>>>
>>>
>>>
>>> On Mon, 10 Sep 2018 at 13:43, Reynold Xin <r...@databricks.com> wrote:
>>>
>>> Well almost all relational databases you can move data in a
>>> transactional way. That’s what transactions are for.
>>>
>>>
>>>
>>> For just straight HDFS, the move is a pretty fast operation so while it
>>> is not completely transactional, the window of potential failure is pretty
>>> short for appends. For writers at the partition level it is fine because it
>>> is just renaming directory, which is atomic.
>>>
>>>
>>>
>>> On Mon, Sep 10, 2018 at 1:40 PM Jungtaek Lim <kabh...@gmail.com> wrote:
>>>
>>> When network partitioning happens it is pretty OK for me to see 2PC not
>>> working, cause we deal with global transaction. Recovery should be hard
>>> thing to get it correctly though. I completely agree it would require
>>> massive changes to Spark.
>>>
>>>
>>>
>>> What I couldn't find for underlying storages is moving data from staging
>>> table to final table in transactional way. I'm not fully sure but as I'm
>>> aware of, many storages would not support moving data, and even HDFS sink
>>> it is not strictly done in transactional way since we move multiple files
>>> with multiple operations. If coordinator just crashes it leaves partial
>>> write, and among writers and coordinator need to deal with ensuring it will
>>> not be going to be duplicated.
>>>
>>>
>>>
>>> Ryan replied me as Iceberg and HBase MVCC timestamps can enable us to
>>> implement "commit" (his reply didn't hit dev. mailing list though) but I'm
>>> not an expert of both twos and I couldn't still imagine it can deal with
>>> various crash cases.
>>>
>>>
>>>
>>> 2018년 9월 11일 (화) 오전 5:17, Reynold Xin <r...@databricks.com>님이 작성:
>>>
>>> I don't think two phase commit would work here at all.
>>>
>>>
>>>
>>> 1. It'd require massive changes to Spark.
>>>
>>>
>>>
>>> 2. Unless the underlying data source can provide an API to coordinate
>>> commits (which few data sources I know provide something like that), 2PC
>>> wouldn't work in the presence of network partitioning. You can't defy the
>>> law of physics.
>>>
>>>
>>>
>>> Really the most common and simple way I've seen this working is through
>>> staging tables and a final transaction to move data from staging table to
>>> final table.
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Mon, Sep 10, 2018 at 12:56 PM Jungtaek Lim <kabh...@gmail.com> wrote:
>>>
>>> I guess we all are aware of limitation of contract on DSv2 writer.
>>> Actually it can be achieved only with HDFS sink (or other filesystem based
>>> sinks) and other external storage are normally not feasible to implement it
>>> because there's no way to couple a transaction with multiple clients as
>>> well as coordinator can't take over transactions from writers to do the
>>> final commit.
>>>
>>>
>>>
>>> XA is also not a trivial one to get it correctly with current execution
>>> model: Spark doesn't require writer tasks to run at the same time but to
>>> achieve 2PC they should run until end of transaction (closing client before
>>> transaction ends normally means aborting transaction). Spark should also
>>> integrate 2PC with its checkpointing mechanism to guarantee completeness of
>>> batch. And it might require different integration for continuous mode.
>>>
>>>
>>>
>>> Jungtaek Lim (HeartSaVioR)
>>>
>>>
>>>
>>> 2018년 9월 11일 (화) 오전 4:37, Arun Mahadevan <ar...@apache.org>님이 작성:
>>>
>>> In some cases the implementations may be ok with eventual consistency
>>> (and does not care if the output is written out atomically)
>>>
>>>
>>>
>>> XA can be one option for datasources that supports it and requires
>>> atomicity but I am not sure how would one implement it with the current
>>> API.
>>>
>>>
>>>
>>> May be we need to discuss improvements at the Datasource V2 API level
>>> (e.g. individual tasks would "prepare" for commit and once the driver
>>> receives "prepared" from all the tasks, a "commit" would be invoked at each
>>> of the individual tasks). Right now the responsibility of the final
>>> "commit" is with the driver and it may not always be possible for the
>>> driver to take over the transactions started by the tasks.
>>>
>>>
>>>
>>>
>>>
>>> On Mon, 10 Sep 2018 at 11:48, Dilip Biswal <dbis...@us.ibm.com> wrote:
>>>
>>> This is a pretty big challenge in general for data sources -- for the
>>> vast majority of data stores, the boundary of a transaction is per client.
>>> That is, you can't have two clients doing writes and coordinating a single
>>> transaction. That's certainly the case for almost all relational databases.
>>> Spark, on the other hand, will have multiple clients (consider each task a
>>> client) writing to the same underlying data store.
>>>
>>>
>>>
>>> DB>> Perhaps we can explore two-phase commit protocol (aka XA) for this
>>> ? Not sure how easy it is to implement this though :-)
>>>
>>>
>>>
>>> Regards,
>>> Dilip Biswal
>>> Tel: 408-463-4980 <(408)%20463-4980>
>>> dbis...@us.ibm.com
>>>
>>>
>>>
>>>
>>>
>>> ----- Original message -----
>>> From: Reynold Xin <r...@databricks.com>
>>> To: Ryan Blue <rb...@netflix.com>
>>> Cc: ross.law...@gmail.com, dev <dev@spark.apache.org>
>>> Subject: Re: DataSourceWriter V2 Api questions
>>> Date: Mon, Sep 10, 2018 10:26 AM
>>>
>>>
>>> I don't think the problem is just whether we have a starting point for
>>> write. As a matter of fact there's always a starting point for write,
>>> whether it is explicit or implicit.
>>>
>>>
>>>
>>> This is a pretty big challenge in general for data sources -- for the
>>> vast majority of data stores, the boundary of a transaction is per client.
>>> That is, you can't have two clients doing writes and coordinating a single
>>> transaction. That's certainly the case for almost all relational databases.
>>> Spark, on the other hand, will have multiple clients (consider each task a
>>> client) writing to the same underlying data store.
>>>
>>>
>>>
>>> On Mon, Sep 10, 2018 at 10:19 AM Ryan Blue <rb...@netflix.com> wrote:
>>>
>>> Ross, I think the intent is to create a single transaction on the
>>> driver, write as part of it in each task, and then commit the transaction
>>> once the tasks complete. Is that possible in your implementation?
>>>
>>>
>>>
>>> I think that part of this is made more difficult by not having a clear
>>> starting point for a write, which we are fixing in the redesign of the v2
>>> API. That will have a method that creates a Write to track the operation.
>>> That can create your transaction when it is created and commit the
>>> transaction when commit is called on it.
>>>
>>>
>>>
>>> rb
>>>
>>>
>>>
>>> On Mon, Sep 10, 2018 at 9:05 AM Reynold Xin <r...@databricks.com> wrote:
>>>
>>> Typically people do it via transactions, or staging tables.
>>>
>>>
>>>
>>>
>>>
>>> On Mon, Sep 10, 2018 at 2:07 AM Ross Lawley <ross.law...@gmail.com>
>>> wrote:
>>>
>>> Hi all,
>>>
>>>
>>>
>>> I've been prototyping an implementation of the DataSource V2 writer for
>>> the MongoDB Spark Connector and I have a couple of questions about how its
>>> intended to be used with database systems. According to the Javadoc for
>>> DataWriter.commit():
>>>
>>>
>>>
>>> *"this method should still "hide" the written data and ask the
>>> DataSourceWriter at driver side to do the final commit via
>>> WriterCommitMessage"*
>>>
>>>
>>>
>>> Although, MongoDB now has transactions, it doesn't have a way to "hide"
>>> the data once it has been written. So as soon as the DataWriter has
>>> committed the data, it has been inserted/updated in the collection and is
>>> discoverable - thereby breaking the documented contract.
>>>
>>>
>>>
>>> I was wondering how other databases systems plan to implement this API
>>> and meet the contract as per the Javadoc?
>>>
>>>
>>>
>>> Many thanks
>>>
>>>
>>>
>>> Ross
>>>
>>>
>>>
>>>
>>>
>>> --
>>>
>>> Ryan Blue
>>>
>>> Software Engineer
>>>
>>> Netflix
>>>
>>>
>>>
>>>
>>> --------------------------------------------------------------------- To
>>> unsubscribe e-mail: dev-unsubscr...@spark.apache.org
>>>
>>>

Reply via email to