Wush Wu
·
·
·
-
-
·
2/38
Ward Cunningham (1992)·
·
3/38
--> -->·
·
·
4/38
API
·
·
·
·
·
·
5/38
·
·
·
6/38
(Sculley et al. 2014)
·
·
·
7/38
8/38
A
B
·
A B ==>
-
·
==>
-
9/38
A B C D ...
A B C D ...
·
·
10/38
11/38
12/38
public/protected/private
IDE
·
·
13/38
14/38
·
·
15/38
16/38
17/38
·
·
18/38
Deep Insight
http://research.google.com/pubs/pub41159.html
19/38
·
·
20/38
·
·
21/38
Automated Feature
Management
Ad Click Prediction: a View from the Trenches·
Chap 8-
Signals consumed by new and active models
are vetted by an automatic system of alerts. Different learn-
ing platforms share a common interface for reporting signal
consumption to a central index. When a signal is depre-
cated (such as when a newer version is made available), we
can quickly identify all consumers of the signal and track
replacement eforts.
22/38
Input Schema
.proto
.proto
jenkins task
·
·
·
23/38
·
-
-
24/38
·
-
-
·
25/38
26/38
Glue
Code
D. Sculley 2014: 5% 95%·
·
Bug
-
-
27/38
LiblineaR,
libffm, ...
·
-
-
28/38
Glue Code
vs
·
·
·
29/38
Pipeline
Jungle
·
·
·
30/38
·
·
31/38
·
·
Knight Capital's
45 4.6
-
32/38
Glue Code Pipeline Jungle·
-
-
-
·
33/38
Version Control Git
API
·
·
Git-
·
D. Sculley: Google-
34/38
·
·
-
35/38
Assertion·
·
Git-
·
36/38
Reference
D. Sculley, et al. (2014) Machine Learning: The High
Interest Credit Card of Technical Debt.
H. Brendan McMahan et al. (2013) Ad Click Prediction: a
View from the Trenches.
·
·
37/38
Q&A
38/38

機器學習的技術債