Streaming Analytics Data analytics recipes for PyToolz
tl;dr: We demonstrate data workflows with Python data structures and PyToolz.
We also introduce
join, a new operation in
In my last two posts I show that Python data structures are
an implementation of
toolz in Cython,
achieves Java speeds on standard Python core data structures like
tuples. As a reminder,
toolz provides functions like
I always give this example when talking about
toolz. It often spurs the
That looks like
GROUP BY from SQL. In what other ways does
toolz let me
do SQL-like operations in Python?
My answer for this is to go look at Pandas which really does a wonderful job
at in-memory data analytics. Toolz targets functional programming more than it
targets data analytics. Still this question is common enough to warrant a
blogpost. The following is my stock answer on how to use pure Python and
cytoolz) for streaming data analytic workflows like selections,
split-apply-combine, and joins. I’ll note throughout when operations are
streaming (can support datasets bigger than memory) or not. This is one of the
few ways in which analysis with
toolz might be preferred over
The toolz functions can be composed to analyze large streaming datasets. Toolz supports common analytics patterns like the selection, grouping, reduction, and joining of data through pure composable functions. These functions often have analogs to familiar operations in other data analytics platforms like SQL or Pandas.
Throughout this post we’ll use this simple dataset of accounts.
Simple projection and linear selection from a sequence is achieved
through the standard functions
These functions correspond to the SQL commands
note: this uses the curried versions of
Of course, these operations are also well supported with standard list/generator comprehension syntax. This syntax is more often used and generally considered to be more Pythonic.
We separate split-apply-combine operations into the following two concepts
- Split the dataset into groups by some property
- Reduce each of the groups with some aggregation function
Toolz supports this common workflow with
- a simple in-memory solution
- a more sophisticated streaming solution.
In Memory Split-Apply-Combine
We first show
valmap separately to show the intermediate
Then we chain them together into a single computation
groupby function collects the entire dataset in memory into a
dictionary. While convenient, the
groupby operation is not streaming
and so this approach is limited to datasets that can fit comfortably
Toolz achieves streaming split-apply-combine with
a function that performs a simultaneous reduction on each group as the
elements stream in. To understand this section you should first be
familiar with the builtin function
reduceby operation takes a key function, like
x, and a binary operator like
lesser = lambda acc, x: acc if acc
< x else x. It successively applies the key function to each item in
succession, accumulating running totals for each key by combining each new
value with the previous total using the binary operator. It can’t accept full
reduction operations like
min as these require access to the entire
group at once. Here is a simple example:
The even numbers are added together
(2 + 4 = 6) into group
the odd numbers are added together
(1 + 3 = 4) into group
Note that we have to replace the reduction
sum with the binary operator
add. The incremental nature of
add allows us to do the summation work as
new data comes in. The use of binary operators like
add over full reductions
sum enables computation on very large streaming datasets.
The challenge to using
reduceby often lies in the construction of a
suitable binary operator. Here is the solution for our accounts example
that adds up the balances for each group:
This construction supports datasets that are much larger than available memory. Only the output must be able to fit comfortably in memory and this is rarely an issue, even for very large split-apply-combine computations.
We register multiple datasets together with join. Consider a second dataset that stores addresses by ID:
We can join this dataset against our accounts dataset by specifying attributes which register different elements with each other; in this case they share a common first column, id.
Join takes four main arguments, a left and right key function and a left
and right sequence. It returns a sequence of pairs of matching items. In our
case the return value of
join is a sequence of pairs of tuples such that the
first element of each tuple (the ID) is the same. In the example above we
unpack this pair of tuples to get the fields that we want (
address) from the result.
Join on arbitrary functions / data
Those familiar with SQL are accustomed to this kind of join on columns. However a functional join is more general than this; it doesn’t need to operate on tuples, and key functions do not need to get particular columns. In the example below we match numbers from two collections so that exactly one is even and one is odd.
The Toolz Join operation fully evaluates the left sequence and streams the right sequence through memory. Thus, if streaming support is desired the larger of the two sequences should always occupy the right side of the join.
The semi-streaming join operation in
toolz is asymptotically optimal.
Computationally it is linear in the size of the input + output. In terms
of storage the left sequence must fit in memory but the right sequence
is free to stream.
The results are not normalized, as in SQL, in that they permit repeated
values. If normalization is desired, consider composing with the
unique (note that
unique is not fully streaming.)
More Complex Example
The accounts example above connects two one-to-one relationships,
addresses; there was exactly one name per ID and one address per ID. This
need not be the case. The join abstraction is sufficiently flexible to join
one-to-many or even many-to-many relationships. The following example finds
city/person pairs where that person has a friend who has a residence in that
city. This is an example of joining two many-to-many relationships because a
person may have many friends and because a friend may have many residences.
Join is computationally powerful:
- It is expressive enough to cover a wide set of analytics operations
- It runs in linear time relative to the size of the input and output
- Only the left sequence must fit in memory
Toolz gives a compact set of primitives for data analysis on plain Python data structures. CyToolz accelerates those workflows through Cython. This approach is both low-tech and supports uncomfortably big data through streaming.
At the same time, Toolz is a general purpose functional standard library, and is not specifically dedicated to data analysis. While there are obvious benefits (streaming, composition, etc.) users interested in data analysis might be better served by using projects dedicated projects like Pandas or SQLAlchemy.
blog comments powered by Disqus