From 04bf93f260b18fe2df8b0c06a0d818b75073582f Mon Sep 17 00:00:00 2001
From: ritchie46 <ritchie46@gmail.com>
Date: Sun, 13 Dec 2020 11:42:19 +0100
Subject: [PATCH] numpy introp

---
 book/src/micro_benchmarks.md | 15 +++++++++++++++
 book/src/numpy.md            | 12 ++++++++++++
 2 files changed, 27 insertions(+)
 create mode 100644 book/src/micro_benchmarks.md
 create mode 100644 book/src/numpy.md

diff --git a/book/src/micro_benchmarks.md b/book/src/micro_benchmarks.md
new file mode 100644
index 000000000..567669a0b
--- /dev/null
+++ b/book/src/micro_benchmarks.md
@@ -0,0 +1,15 @@
+# Micro benchmarks
+Below are some micro benchmarks shown between Polars and Pandas. Note that these are just micro benchmarks and nothing
+more. An optimization can lead to increased performance in a single benchmark and lead to regressions somewhere else.
+
+To truly make performance comparisons we should at least look at the macro level of a query. 
+
+## Csv parsing
+![](../img/csv.png)
+
+# Joins
+![](../img/join_80_000.png)
+
+## Groupby
+![](../img/groupby10_.png)
+![](../img/groupby10_mem.png)
diff --git a/book/src/numpy.md b/book/src/numpy.md
new file mode 100644
index 000000000..290d36418
--- /dev/null
+++ b/book/src/numpy.md
@@ -0,0 +1,12 @@
+# Numpy interoperability
+
+Polars Series have support for numpy's [universal functions](https://numpy.org/doc/stable/reference/ufuncs.html).
+That means that numpys elementwise function like `np.exp`, `np.cos`, `np.div`, etc. all work with almost zero overhead.
+There are few gotcha's however. Missing values are ignored during the function application. They are maintained, as they
+are just a separate bitmask. However, any function that depends on previous/next elements, i.e. a sum, a convolve etc.
+should not be used or used with caution.
+
+## Conversion
+You can convert a `Series` to a numpy array with the `.to_numpy` method. Missing values will be replaced by `NaN` during
+the conversion. If the Series doesn't have missing values, or you don't care about them, you can use the `.view` method.
+This provides a zero copy numpy array of the `Series` data.