spark-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Katherine Prevost <...@hypatian.org>
Subject Re: Questions about the future of UDTs and Encoders
Date Wed, 16 Aug 2017 15:55:23 GMT
I'd say the quick summary of the problem is this:

The encoder mechanism does not deal well with fields of case classes (you
must use builtin types (including other case classes) for case class
fields), and UDTs are not currently available (and never integrated well
with built-in operations).

Encoders work great for individual fields if you're using tuples, but once
you get up over four or five fields this becomes incomprehensible. And, of
course, encoders do nothing for you once you are in the realm of dataframes
(including operations on fields, results of dataframe-based methods, and
working in languages other than Scala.)

The sort of machinations I describe below are unpleasant but not a huge
deal for people who are trained as developers... but they're a much bigger
mess when we have to provide these interfaces to our data scientists. Yes,
they can do it, but the "every address is a string and you have to use
these functions that parse the strings over and over again" approach is
easier to use (if massively inefficient).

I would like to improve Spark so that we can provide these types that our
data scientists need to use *all the time* in a way that's both efficient
and easy to use.

Hence, my interest in doing work on the UDT and/or Encoder mechanisms of
Spark (or equivalent, if something new is in the works), and my interest in
hearing from anybody who is already working in this area, or hearing about
any future plans that have already been made in this area.


In more detail:

On Wed, Aug 16, 2017 at 2:49 AM Jörn Franke <jornfranke@gmail.com> wrote:

> Not sure I got to fully understand the issue (source code is always
> helpful ;-) but why don't you override the toString method of IPAddress.
> So, IP address could still be byte , but when it is displayed then toString
> converts the byteaddress into something human-readable?
>

There are a couple of reasons it's not that simple. (If you look at the
sample snippets of code I did include, you'll see that I did define
toString methods.)

The first problem is basically because toString doesn't happen when working
with DataFrames, which are often the result of common Spark operations in
Scala (though staying in the realm of Datasets is getting easier, and
apparently also becoming more efficient). Outside of Scala, it's DataFrames
all the way down.

(If you look at my example code, you'll also see what happens when you have
a DataFrame with a field that is a struct with a byte array in it, and
nobody ever wants to see "[B@617f4814".)

You can get around that (as long as you're still in a Dataset) with
something like this (this is using the IPAddress.toString method to produce
"IPAddress(Array(1,2,3,4))"):

scala> ys.take(20)
res10: Array[Rec] = Array(Rec(IPAddress(Array(1, 2, 3, 4)),
IPAddress(Array(5, 6, 7, 8))), Rec(IPAddress(Array(1, 2, 3, 4, 5, 6, 7, 8,
9, 10, 11, 12, 13, 14, 15, 16)), IPAddress(Array(17, 18, 19, 20, 21, 22,
23, 24, 25, 26, 27, 28, 29, 30, 31, 32))))

But then of course you lose any easy ability to view Rec fields in columns.
(And while you could make something that prints Rec as columns, what
happens once you transform your record and turn it into a tuple?)

The second one is that operating on the fields cleanly is still rather
painful, even if the values were to be displayed cleanly. This is what you
have to do to search for rows that have a specific IPAddress value (ys("a")
is a column of IPAddress, a is an IPAddress):

scala> ys.select(ys("a.bytes") === a.bytes)
res9: org.apache.spark.sql.DataFrame = [(a.bytes AS `bytes` = X'01020304'):
boolean]

It's worth noting that an implicit conversion from IPAddress to Array[Byte]
or to Column wouldn't work here, because === accepts Any.


katherine.

> On 15. Aug 2017, at 18:49, Katherine Prevost <prevost1@cert.org> wrote:
> >
> > Hi, all!
> >
> >
> > I'm a developer who works to support data scientists at CERT. We've
> > been having some great success working with Spark for data analysis,
> > and I have some questions about how we could contribute to work on
> > Spark in support of our goals.
> >
> > Specifically, we have some interest in user-defined types, or their
> > equivalents.
> >
> >
> > When Spark 2 arrived, user-defined types (UDTs) were made private and
> > seem to have fallen by the wayside in favor of using encoders for
> > Datasets. I have some questions about the future of these mechanisms,
> > and was wondering if there's been a plan published anywhere for the
> > future of these mechanisms, or anyone I could talk to about where
> > things are going with them.
> >
> > I've roughly outlined our experience with these two mechanisms below,
> > and our hopes for what might be accomplished in the future.
> >
> > We'd love to spend some effort on development here, but haven't been
> > able to figure out if anyone is already working on improvements in
> > this area, or if there's some plan in place for where things are going
> > to go.
> >
> > So, I'd love to get in touch with anyone who might know more.
> >
> >
> > Background:
> >
> > Much of the work in my group is analysis of Internet protocol data,
> > and I think that IP addresses are a great example how a custom atomic
> > type can be helpful.
> >
> > IP addresses (including both 32-bit IPv4 addresses and 128-bit IPv6
> > addresses) have a natural binary form (a sequence of bytes). Using
> > this format makes the default implementation of certain basic
> > operations sensible (equality and comparison, for example). Defining
> > UDFs for more complicated operations is not terribly difficultt. But
> > this format is not human-friendly to view.
> >
> > The human-readable presentations of IP addresses, on the other hand,
> > are large and unwieldy to work with computationally. There is a
> > canonical textual form for both IPv4 and IPv6 addresses, but
> > converting back and forth between that form and the binary form is
> > expensive, and the text form is generally at least twice as large as
> > the binary form. The text form is suitable for presenting to human
> > beings, but that's about it.
> >
> > There are also a variety of other types of Internet data that are best
> > represented by byte arrays and the like, meaning that simply saying
> > "just use a byte array for this column!" can be unfortunate for both
> > type-safety and comprehensibility of a colleciton of data.
> >
> >
> > When we were working on top of Spark 1, we had begun to look using
> > UDTs to represent IP addresses. There were some issues with working
> > with UDTs and working with the built-in operations like comparisons,
> > but we had some hope for improvements with future Spark releases.
> >
> > With Spark 2.0, the UDT API was made private, and the encoder
> > mechanism was suggested for use instead. For a bit, we experimented
> > with using the API anyway by putting stubs into Spark's namespace, but
> > there weren't really a lot of good places to hook various operations
> > like equality that one would expect to work on an atomic type.
> >
> >
> > We also tried using the encoder APIs, and ran into a few problems
> > there as well. Encoders are well suited to handling "top-level"
> > values, but the most convenient way to work with encoded data is by
> > having your top level be a case class defining types and names for a
> > record type. And here, there's a problem, because encoders from the
> > implicit environment are not available when encoding the fields of a
> > case class. So, if we defined a custom encoder for our IPAddress type,
> > and then included an IPAddress as a field of a record, this would
> > result in an error.
> >
> > One approach we tried to get around that was to make IP addresses
> > themselves into case classes as well, so that only the default
> > encoders would be required. This eliminated the error, but made
> > working with the values a nightmare. If we made a Dataset[IPAddress],
> > the byte array would be presented in a reasonable manner, but a
> > Dataset[Rec] where Rec had IPAddress fields was another story,
> > resulting in the default toString of Java arrays being used:
> >
> > +-------------+-------------+
> > |            a|            b|
> > +-------------+-------------+
> > |[[B@47260109]|[[B@3538740a]|
> > |[[B@617f4814]|[[B@77e69bee]|
> > +-------------+-------------+
> >
> > (See code snippet at the end of this message for details.)
> >
> > Now basically all interactions would have to go through UDFs,
> > including remembering to format the IPAddress field if you wanted any
> > useful information out of it at all.
> >
> >
> > As a result, since our initial experiments with 2.0 we dropped back
> > and punted to using text for all IP addresses. But, we'd still like to
> > do better. What we optimally want is some mechanism for a user-defined
> > atomic type (whether via encoders or via registering a new type) which
> > allows for:
> >
> > * An appropriately efficient underlying form to be used. (A struct
> >   with a byte array field would be fine. A byte array field would be
> >   fine.)
> >
> > * A display form that is meaningful to the user (the expected form
> >   like "172.217.5.238" and "2607:f8b0:4004:800::200e".)
> >
> > * At least some support for standard SQL operators like equality and
> >   comparison, and the ability to define UDFs that work with the type.
> >
> > Longer term, it would be lovely to:
> >
> > * Be able to work with values of the type in an appropriate way in
> >   different source languags (i.e. make it not hard to work with the
> >   values in Python or R, although the restrictions of those languages
> >   will require additional implementation work.)
> >
> > * Be able to provide new Catalyst optimizations specific to the type
> >   and functions defined on the type.
> >
> > We'd love to provide some effort at achieving these goals, but aren't
> > sure where to start. We'd like to avoid stepping in the way of any
> > efforts that might already be underway to improve these mechanisms.
> >
> >
> > Thanks very much!
> >
> > Katherine Prevost
> > Carnegie Mellon / Software Engineering Institute / CERT
> >
> >
> > -------------------------------------------------------------------->8--
> >
> > // Simple example demonstrating the treatment of a case class with a
> > // byte array within another case class.
> >
> > case class IPAddress(bytes: Array[Byte]) {
> >  override def toString: String = s"IPAddress(Array(${bytes.mkString(",
> ")}))"
> > }
> >
> > val a = IPAddress(Array(1,2,3,4))
> > val b = IPAddress(Array(5,6,7,8))
> > val c = IPAddress(Array(1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16))
> > val d = IPAddress(Array(17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32))
> >
> > val x = Array(a, b, c, d)
> > val xs = sc.parallelize(x).toDS
> >
> > /*
> > scala> xs.show
> > +--------------------+
> > |               bytes|
> > +--------------------+
> > |       [01 02 03 04]|
> > |       [05 06 07 08]|
> > |[01 02 03 04 05 0...|
> > |[11 12 13 14 15 1...|
> > +--------------------+
> > */
> >
> > case class Rec(a: IPAddress, b: IPAddress) {
> >  override def toString: String = s"Rec($a, $b)"
> > }
> >
> > val e = Rec(a, b)
> > val f = Rec(c, d)
> > val y = Array(e, f)
> > val ys = sc.parallelize(y).toDS
> >
> > /*
> > scala> ys.show
> > +-------------+-------------+
> > |            a|            b|
> > +-------------+-------------+
> > |[[B@47260109]|[[B@3538740a]|
> > |[[B@617f4814]|[[B@77e69bee]|
> > +-------------+-------------+
> > */
> >
> > -------------------------------------------------------------------->8--
> >
> > ---------------------------------------------------------------------
> > To unsubscribe e-mail: dev-unsubscribe@spark.apache.org
> >
>
> ---------------------------------------------------------------------
> To unsubscribe e-mail: dev-unsubscribe@spark.apache.org
>
>

Mime
View raw message