ETL for Medium Data

Four distinct ideas/challenges

1. A Grammar for
2. Reproducible and Painless
3. Extract-Transform-Load Operations
4. on Medium Data

An example: Citi Bike

Citi Bike: problems?

Citi Bike: research

• Data Analysis and optimization for (Citi)Bike sharing [O’Mahony, Shmoys (2015)]
• Predicting bike usage for New York City’s bike sharing system [Singhvi, et al. (2015)]
• Smarter tools for (Citi) bike sharing [O’Mahony (2015)]
• Incorporating the impact of spatio-temporal interactions on bicycle sharing system demand: A case study of New York CitiBike system [Faghih-Imani, Eluru (2016)]

Citi Bike: data

We obtained bike usage statistics for April, May, June and July 2014 from Citi Bike’s website (https://www.citibikenyc.com/system-data). This dataset contains start station id, end station id, station latitude, station longitude and trip time for each bike trip. 332 bike stations have one or more originating bike trips. 253 of these are in Manhattan while 79 are in Brooklyn (left panel of Figure 1). We processed this raw data to get the number of bike trips between each station pair during morning rush hours.
–Singhvi, et al. (2015)

• Could you reproduce this analysis?

Citi Bike: set up citibike database

• Set up bikes object

library(citibike)

bikes <- etl("citibike", 
             dir = "~/dumps/citibike/",
             db = src_mysql_cnf("citibike"))

• Populate a database

bikes %>%
  etl_update(years = 2014, months = 4:7)

Citi Bike: query database

trips <- bikes %>%
  tbl("trips")

trips %>%
  group_by(Start_Station_ID) %>%
  summarize(num_trips = n()) %>%
  filter(num_trips >= 1) %>%
  arrange(desc(num_trips)) %>%
  collect()

## # A tibble: 332 x 2
##    Start_Station_ID num_trips
##               <int>     <dbl>
##  1              519     50316
##  2              521     49511
##  3              293     45391
##  4              497     45154
##  5              426     40046
##  6              435     37542
##  7              285     36074
##  8              499     33849
##  9              151     33776
## 10              444     33663
## # ... with 322 more rows

Citi Bike: reproducibility?

• How confident are you that we have a copy of the same data as these researchers?

• same number of stations…
  • same number of rows?
  • rows contain the same information?
  • integer overflows?
• impossible to verify given the description

Replicability Crisis

• Why Most Published Research Findings Are False (Ioannidis, 2005)
• Announcement: Reducing our irreproducibility (Nature, 2013)
• ASA’s Statement on p-Values (Wasserstein & Lazar, 2016)

Reproducible scholarship

An article about a computational result is advertising, not scholarship. The actual scholarship is the full software environment, code and data, that pro- duced the result.
–Donoho (2010) paraphrasing Claerbout (1994)

• Must include the code you wrote to get the data!

Data size for the single user

ETL operations

• Extract -> download
• Transform -> wrangle
• Load -> into database (typically SQL)

A grammar?

In linguistics, grammar is the set of structural rules governing the composition of clauses, phrases, and words in any given natural language.
–Wikipedia

• The Grammar of Graphics (Wilkinson, 2006)
  • ggplot2 (Wickham, 2009)
• dplyr: a grammar of data manipulation (Wickham & Francois, 2016)
• Covers many use cases with relatively few words
• Once you learn the grammar, you can speak freely

A grammar for ETL

• nouns:
  • objects that inherit from class etl
• verbs:
  • etl_extract()
  • etl_transform()
  • etl_load()
• adverbs:
  • arguments to “verb” functions
• predictable, pipeable syntax

ETL suite of packages

• A CRAN package that provides a framework for ETL
  • etl  
• An ecosystem of dependent packages for other data sources
  • macleish  
  • airlines  
  • imdb  
  • nyctaxi  
  • nyc311  
  • fec  
  • citibike  
  • roll your own!

Example: airlines

• Create an empty database in MySQL

system("mysql -e 'CREATE DATABASE IF NOT EXISTS airlines;'")

• Set up the etl object

library(airlines)

src_db <- src_mysql_cnf("airlines")

ontime <- etl("airlines", db = src_db, dir = "~/dumps/airlines") 

Example: airlines cont’d

• Perform ETL operations

ontime %>%
  etl_extract(years = 1987:2017) %>%
  etl_transform(years = 2001:2010) %>%
  etl_load(years = 2009:2010, months = c(1:3, 5))

• May take hours!

Example: airlines cont’d

• Q: Which airline had the shortest average delay at SEA?

ontime %>%
  tbl("flights") %>%
  filter(year == 2010, dest == "SEA") %>%
  group_by(carrier) %>%
  summarize(num_flights = n(), 
            begin = min(month), end = max(month), 
            avg_delay = mean(arr_delay, na.rm = TRUE)) %>%
  arrange(avg_delay)

## # Source:     lazy query [?? x 5]
## # Database:   mysql 5.7.22-0ubuntu0.16.04.1 [root@127.0.0.1:/airlines]
## # Ordered by: avg_delay
##    carrier num_flights begin   end avg_delay
##    <chr>         <dbl> <int> <int>     <dbl>
##  1 UA             2236     1     5  -11.4   
##  2 DL             2598     1     5  -10.3   
##  3 US              941     1     5   -7.61  
##  4 AS            13488     1     5   -5.83  
##  5 B6              498     1     5   -3.39  
##  6 FL              196     1     5   -3.17  
##  7 OO             2747     1     5   -2.08  
##  8 WN             3919     1     5   -2.05  
##  9 AA             1591     1     5   -1.09  
## 10 HA              295     1     5   -0.0949
## # ... with more rows

Noun: an etl object is

• a dplyr::src_sql object
  • provides a connection to a SQL database
  • local or remote
  • SQLite by default (tempfile)
  • Any dbplyr backend supported
  • uses DBI::dbWriteTable methods

class(cars)

## [1] "etl_mtcars" "etl"        "src_dbi"    "src_sql"    "src"

Noun: an etl object has

• a place to store files
  • local storage for raw and processed data
  • tempdir by default

summary(ontime)

## files:
##    n     size                              path
## 1 13 0.193 GB  /home/bbaumer/dumps/airlines/raw
## 2 11  0.49 GB /home/bbaumer/dumps/airlines/load

##       Length Class           Mode       
## con   1      MySQLConnection S4         
## disco 2      -none-          environment

Verbs: chaining operations

getS3method("etl_update", "default")

## function(obj, ...) {
##   obj <- obj %>%
##     etl_extract(...) %>%
##     etl_transform(...) %>%
##     etl_load(...)
##   invisible(obj)
## }
## <environment: namespace:etl>

Extending etl

The End

• Thanks to the dplyr and rstats-db developers!!
• Read the paper (https://arxiv.org/abs/1708.07073)
  • Coming soon to Journal of Computational and Graphical Statistics!
• Check out these slides: (https://bit.ly/2LzLZtm)
  
  

beanumber     @BaumerBen