Wes et al, I completed a preliminary study of populating a Feather file
incrementally. Some notes and questions:
I wrote the following dataframe to a feather file:
a b
0 0123456789 0.0
1 0123456789 NaN
2 0123456789 NaN
3 0123456789 NaN
4 None NaN
In re-writing the flatbuffers metadata (flatc -p doesn't
support --gen-mutable! yuck! C++ to the rescue...), it seems that
read_feather is not affected by NumRows? It seems to be driven entirely by
the per-column Length values?
Also, it seems as if one of the primary usages of NullCount is to determine
whether or not a bitfield is present? In the initialization data above I
was careful to have a null value in each column in order to generate a
bitfield.
I then wiped the bitfields in the file and set all of the string indices to
one past the end of the blob buffer (all strings empty):
a b
0 None NaN
1 None NaN
2 None NaN
3 None NaN
4 None NaN
I then set the first record to some data by consuming some of the string
blob and row 0 and 1 indices, also setting the double:
a b
0 Hello, world! 5.0
1 None NaN
2 None NaN
3 None NaN
4 None NaN
As mentioned above, NumRows seems to be ignored. I tried adjusting each
column Length to mask off higher rows and it worked for 4 (hide last row)
but not for less than 4. I take this to be due to math used to figure out
where the buffers are relative to one another since there is only one
metadata offset for all of: the (optional) bitset, index column and (if
string) blobs.
Populating subsequent rows would proceed in a similar way until all of the
blob storage has been consumed, which may come before the pre-allocated
rows have been consumed.
So what does this mean for my desire to incrementally write these
(potentially memory-mapped) pre-allocated files and/or Arrow buffers in
memory? Some thoughts:
- If a single value (such as NumRows) were consulted to determine the table
row dimension then updating this single value would serve to tell a reader
which rows are relevant. Assuming this value is updated after all other
mutations are complete, and assuming that mutations are only appends
(addition of rows), then concurrency control involves only ensuring that
this value is not examined while it is being written.
- NullCount presents a concurrency problem if someone reads the file after
this field has been updated, but before NumRows has exposed the new record
(or vice versa). The idea previously mentioned that there will "likely
[be] more statistics in the future" feels like it might be scope creep to
me? This is a data representation, not a calculation framework? If
NullCount had its genesis in the optional nature of the bitfield, I would
suggest that perhaps NullCount can be dropped in favor of always supplying
the bitfield, which in any event is already contemplated by the spec:
"Implementations may choose to always allocate one anyway as a matter of
convenience." If the concern is space savings, Arrow is already an
extremely uncompressed format. (Compression is something I would also
consider to be scope creep as regards Feather... compressed filesystems can
be employed and there are other compressed dataframe formats.) However, if
protecting the 4 bytes required to update NowRows turns out to be no easier
than protecting all of the statistical bytes as well as part of the same
"critical section" (locks: yuck!!) then statistics pose no issue. I have a
feeling that the availability of an atomic write of 4 bytes will depend on
the storage mechanism... memory vs memory map vs write() etc.
- The elephant in the room appears to be the presumptive binary yes/no on
mutability of Arrow buffers. Perhaps the thought is that certain batch
processes will be wrecked if anyone anywhere is mutating buffers in any
way. But keep in mind I am not proposing general mutability, only
appending of new data. *A huge amount of batch processing that will take
place with Arrow is on time-series data (whether financial or otherwise).
It is only natural that architects will want the minimal impedance mismatch
when it comes time to grow those time series as the events occur going
forward.* So rather than say that I want "mutable" Arrow buffers, I would
pitch this as a call for "immutable populated areas" of Arrow buffers
combined with the concept that the populated area can grow up to whatever
was preallocated. This will not affect anyone who has "memoized" a
dimension and wants to continue to consider the then-current data as
immutable... it will be immutable and will always be immutable according to
that then-current dimension.
Thanks in advance for considering this feedback! I absolutely require
efficient row-wise growth of an Arrow-like buffer to deal with time series
data in near real time. I believe that preallocation is (by far) the most
efficient way to accomplish this. I hope to be able to use Arrow! If I
cannot use Arrow than I will be using a home-grown Arrow that is identical
except for this feature, which would be very sad! Even if Arrow itself
could be used in this manner today, I would be hesitant to use it if the
use-case was not protected on a go-forward basis.
Of course, I am completely open to alternative ideas and approaches!
-John
On Mon, May 6, 2019 at 11:39 AM Wes McKinney <wesmck...@gmail.com> wrote:
> hi John -- again, I would caution you against using Feather files for
> issues of longevity -- the internal memory layout of those files is a
> "dead man walking" so to speak.
>
> I would advise against forking the project, IMHO that is a dark path
> that leads nowhere good. We have a large community here and we accept
> pull requests -- I think the challenge is going to be defining the use
> case to suitable clarity that a general purpose solution can be
> developed.
>
> - Wes
>
>
> On Mon, May 6, 2019 at 11:16 AM John Muehlhausen <j...@jgm.org> wrote:
> >
> > François, Wes,
> >
> > Thanks for the feedback. I think the most practical thing for me to do
> is
> > 1- write a Feather file that is structured to pre-allocate the space I
> need
> > (e.g. initial variable-length strings are of average size)
> > 2- come up with code to monkey around with the values contained in the
> > vectors so that before and after each manipulation the file is valid as I
> > walk the rows ... this is a writer that uses memory mapping
> > 3- check back in here once that works, assuming that it does, to see if
> we
> > can bless certain mutation paths
> > 4- if we can't bless certain mutation paths, fork the project for those
> who
> > care more about stream processing? That would not seem to be ideal as I
> > think mutation in row-order across the data set is relatively low impact
> on
> > the overall design?
> >
> > Thanks again for engaging the topic!
> > -John
> >
> > On Mon, May 6, 2019 at 10:26 AM Francois Saint-Jacques <
> > fsaintjacq...@gmail.com> wrote:
> >
> > > Hello John,
> > >
> > > Arrow is not yet suited for partial writes. The specification only
> > > talks about fully frozen/immutable objects, you're in implementation
> > > defined territory here. For example, the C++ library assumes the Array
> > > object is immutable; it memoize the null count, and likely more
> > > statistics in the future.
> > >
> > > If you want to use pre-allocated buffers and array, you can use the
> > > column validity bitmap for this purpose, e.g. set all null by default
> > > and flip once the row is written. It suffers from concurrency issues
> > > (+ invalidation issues as pointed) when dealing with multiple columns.
> > > You'll have to use a barrier of some kind, e.g. a per-batch global
> > > atomic (if append-only), or dedicated column(s) à-la MVCC. But then,
> > > the reader needs to be aware of this and compute a mask each time it
> > > needs to query the partial batch.
> > >
> > > This is a common columnar database problem, see [1] for a recent paper
> > > on the subject. The usual technique is to store the recent data
> > > row-wise, and transform it in column-wise when a threshold is met akin
> > > to a compaction phase. There was a somewhat related thread [2] lately
> > > about streaming vs batching. In the end, I think your solution will be
> > > very application specific.
> > >
> > > François
> > >
> > > [1] https://db.in.tum.de/downloads/publications/datablocks.pdf
> > > [2]
> > >
> https://lists.apache.org/thread.html/27945533db782361143586fd77ca08e15e96e2f2a5250ff084b462d6@%3Cdev.arrow.apache.org%3E
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > > On Mon, May 6, 2019 at 10:39 AM John Muehlhausen <j...@jgm.org> wrote:
> > > >
> > > > Wes,
> > > >
> > > > I’m not afraid of writing my own C++ code to deal with all of this
> on the
> > > > writer side. I just need a way to “append” (incrementally populate)
> e.g.
> > > > feather files so that a person using e.g. pyarrow doesn’t suffer some
> > > > catastrophic failure... and “on the side” I tell them which rows are
> junk
> > > > and deal with any concurrency issues that can’t be solved in the
> arena of
> > > > atomicity and ordering of ops. For now I care about basic types but
> > > > including variable-width strings.
> > > >
> > > > For event-processing, I think Arrow has to have the concept of a
> > > partially
> > > > full record set. Some alternatives are:
> > > > - have a batch size of one, thus littering the landscape with
> trivially
> > > > small Arrow buffers
> > > > - artificially increase latency with a batch size larger than one,
> but
> > > not
> > > > processing any data until a batch is complete
> > > > - continuously re-write the (entire!) “main” buffer as batches of
> length
> > > 1
> > > > roll in
> > > > - instead of one main buffer, several, and at some threshold combine
> the
> > > > last N length-1 batches into a length N buffer ... still an
> inefficiency
> > > >
> > > > Consider the case of QAbstractTableModel as the underlying data for a
> > > table
> > > > or a chart. This visualization shows all of the data for the recent
> past
> > > > as well as events rolling in. If this model interface is
> implemented as
> > > a
> > > > view onto “many thousands” of individual event buffers then we gain
> > > nothing
> > > > from columnar layout. (Suppose there are tons of columns and most of
> > > them
> > > > are scrolled out of the view.). Likewise we cannot re-write the
> entire
> > > > model on each event... time complexity blows up. What we want is to
> > > have a
> > > > large pre-allocated chunk and just change rowCount() as data is
> > > “appended.”
> > > > Sure, we may run out of space and have another and another chunk for
> > > > future row ranges, but a handful of chunks chained together is better
> > > than
> > > > as many chunks as there were events!
> > > >
> > > > And again, having a batch size >1 and delaying the data until a
> batch is
> > > > full is a non-starter.
> > > >
> > > > I am really hoping to see partially-filled buffers as something we
> keep
> > > our
> > > > finger on moving forward! Or else, what am I missing?
> > > >
> > > > -John
> > > >
> > > > On Mon, May 6, 2019 at 8:24 AM Wes McKinney <wesmck...@gmail.com>
> wrote:
> > > >
> > > > > hi John,
> > > > >
> > > > > In C++ the builder classes don't yet support writing into
> preallocated
> > > > > memory. It would be tricky for applications to determine a priori
> > > > > which segments of memory to pass to the builder. It seems only
> > > > > feasible for primitive / fixed-size types so my guess would be
> that a
> > > > > separate set of interfaces would need to be developed for this
> task.
> > > > >
> > > > > - Wes
> > > > >
> > > > > On Mon, May 6, 2019 at 8:18 AM Jacques Nadeau <jacq...@apache.org>
> > > wrote:
> > > > > >
> > > > > > This is more of a question of implementation versus
> specification. An
> > > > > arrow
> > > > > > buffer is generally built and then sealed. In different
> languages,
> > > this
> > > > > > building process works differently (a concern of the language
> rather
> > > than
> > > > > > the memory specification). We don't currently allow a half built
> > > vector
> > > > > to
> > > > > > be moved to another language and then be further built. So the
> > > question
> > > > > is
> > > > > > really more concrete: what language are you looking at and what
> is
> > > the
> > > > > > specific pattern you're trying to undertake for building.
> > > > > >
> > > > > > If you're trying to go across independent processes (whether the
> same
> > > > > > process restarted or two separate processes active
> simultaneously)
> > > you'll
> > > > > > need to build up your own data structures to help with this.
> > > > > >
> > > > > > On Mon, May 6, 2019 at 6:28 PM John Muehlhausen <j...@jgm.org>
> wrote:
> > > > > >
> > > > > > > Hello,
> > > > > > >
> > > > > > > Glad to learn of this project— good work!
> > > > > > >
> > > > > > > If I allocate a single chunk of memory and start building Arrow
> > > format
> > > > > > > within it, does this chunk save any state regarding my
> progress?
> > > > > > >
> > > > > > > For example, suppose I allocate a column for floating point
> (fixed
> > > > > width)
> > > > > > > and a column for string (variable width). Suppose I start
> > > building the
> > > > > > > floating point column at offset X into my single buffer, and
> the
> > > string
> > > > > > > “pointer” column at offset Y into the same single buffer, and
> the
> > > > > string
> > > > > > > data elements at offset Z.
> > > > > > >
> > > > > > > I write one floating point number and one string, then go away.
> > > When I
> > > > > > > come back to this buffer to append another value, does the
> buffer
> > > > > itself
> > > > > > > know where I would begin? I.e. is there a differentiation in
> the
> > > > > column
> > > > > > > (or blob) data itself between the available space and the used
> > > space?
> > > > > > >
> > > > > > > Suppose I write a lot of large variable width strings and “run
> > > out” of
> > > > > > > space for them before running out of space for floating point
> > > numbers
> > > > > or
> > > > > > > string pointers. (I guessed badly when doing the original
> > > > > allocation.). I
> > > > > > > consider this to be Ok since I can always “copy” the data to
> > > “compress
> > > > > out”
> > > > > > > the unused fp/pointer buckets... the choice is up to me.
> > > > > > >
> > > > > > > The above applied to a (feather?) file is how I anticipate
> > > appending
> > > > > data
> > > > > > > to disk... pre-allocate a mem-mapped file and gradually fill
> it up.
> > > > > The
> > > > > > > efficiency of file utilization will depend on my projections
> > > regarding
> > > > > > > variable-width data types, but as I said above, I can always
> > > re-write
> > > > > the
> > > > > > > file if/when this bothers me.
> > > > > > >
> > > > > > > Is this the recommended and supported approach for incremental
> > > appends?
> > > > > > > I’m really hoping to use Arrow instead of rolling my own, but
> > > > > functionality
> > > > > > > like this is absolutely key! Hoping not to use a side-car
> file (or
> > > > > memory
> > > > > > > chunk) to store “append progress” information.
> > > > > > >
> > > > > > > I am brand new to this project so please forgive me if I have
> > > > > overlooked
> > > > > > > something obvious. And again, looks like great work so far!
> > > > > > >
> > > > > > > Thanks!
> > > > > > > -John
> > > > > > >
> > > > >
> > >
>
