ℹ️ NOTE: This tour is primarily targeted to Linux and macOS users. Though qsv works on Windows, the tour assumes basic knowledge of command-line piping and redirection, and uses other command-line tools (curl, tee, head, etc.) that are not installed by default on Windows.
For a more detailed, interactive tour (which also happens to be Windows-friendly) see 100.dathere.com.
Let's say you're playing with some data from the Data Science Toolkit, which contains several CSV files. Maybe you're interested in the population counts of each city in the world. So grab the 124MB, 2.7M row CSV file and start examining it:
# there are no headers in the original repo, so let's download a prepared CSV with headers
$ curl -LO https://raw.githubusercontent.com/wiki/dathere/qsv/files/wcp.zip
$ unzip wcp.zip
$ qsv headers wcp.csv
1 Country
2 City
3 AccentCity
4 Region
5 Population
6 Latitude
7 Longitude
The next thing you might want to do is get an overview of the kind of data that
appears in each column. The stats command will do this for you:
$ qsv stats wcp.csv | qsv table
field type sum min max min_length max_length mean stddev variance nullcount
Country String ad zw 2 2 0
City String al lusayli ??ykkvibaer 1 87 0
AccentCity String Al Lusayli ??zl??ce 1 87 0
Region String 00 Z4 0 2 4
Population Integer 2290536128 3 31480498 0 8 48729.62723114559 308410.84307353816 95117248125.33058 2652349
Latitude Float 76585211.1977638 -54.9333333 82.483333 1 12 28.371681223642454 21.938373536961045 481.2922334472327 0
Longitude Float 75976506.66428813 -179.9833333 180.0 1 14 28.14618114715136 62.472858625866586 3902.858064887513 0
Wow! That was fast! It took just 1.3 seconds to compile all that.1 One reason for qsv's speed is that it mainly works in "streaming" mode - computing statistics as it "streams" the CSV file line by line. This also means it can gather statistics on arbitrarily large files, as it does not have to load the entire file into memory.2
But can we get more summary statistics? What's the variance, the modes, the distribution (quartiles),
and the cardinality of the data? No problem. That's why qsv stats has an --everything option to
compute these more "expensive" stats. Expensive - as these extended statistics can only be computed at
the cost of loading the entire file into memory.
$ qsv stats wcp.csv --everything | qsv table
field type sum min max min_length max_length mean stddev variance nullcount lower_outer_fence lower_inner_fence q1 q2_median q3 iqr upper_inner_fence upper_outer_fence skewness mode cardinality
Country String ad zw 2 2 0 ru 231
City String al lusayli ??ykkvibaer 1 87 0 san jose 2008182
AccentCity String Al Lusayli ??zl??ce 1 87 0 San Antonio 2031214
Region String 00 Z4 0 2 4 04 392
Population Integer 2290536128 3 31480498 0 8 48729.627231145605 308410.84307353816 95117248125.33058 2652349 -69768.5 -33019.25 3730.0 10879.0 28229.5 24499.5 64978.75 101728.0 0.4163962529847548 28461
Latitude Float 76585211.1977638 -54.9333333 82.483333 1 12 28.37168122364246 21.938373536961045 481.2922334472327 0 -84.7705556 -35.9076389 12.9552778 33.8666667 45.5305556 32.5752778 94.3934723 143.256389 -0.28388092518431285 50.8 255133
Longitude Float 75976506.66428813 -179.9833333 180.0 1 14 28.146181147151353 62.472858625866586 3902.858064887513 0 -199.36666790000004 -98.49166745000002 2.383333 26.8802778 69.6333333 67.25000030000001 170.50833375000002 271.38333420000004 0.2714663289005219 23.1 407568
ℹ️ NOTE: The
qsv tablecommand takes any CSV data and formats it into aligned columns using elastic tabstops. You'll notice that it even gets alignment right with respect to Unicode characters.
So, this command took 3.22 seconds to run on my machine, but we can speed it up by creating an index and re-running the command:
qsv index wcp.csv
qsv stats wcp.csv --everything | qsv table
Which cuts it down to 1.95 seconds - 1.65x faster! (And creating the 21.6mb index took 0.27 seconds.
What about the first stats without --everything? From 1.3 seconds to 0.16 seconds with an index - 8.25x faster!)
Notably, the same type of "statistics" command in another
CSV command line toolkit
takes about 10 seconds to produce a subset of statistics on the same data set. Visidata
takes much longer - ~1.5 minutes to calculate a subset of these statistics with its Describe sheet.
Even python pandas'
describe(include="all")) took 12 seconds to calculate a subset of qsv's "streaming" statistics.3
This is another reason for qsv's speed. Creating an index accelerated statistics gathering as it enables multithreading & fast I/O.
For multithreading - running stats with an index was 8.25x faster because it divided the file into
16 equal chunks1 with ~170k records each, then running stats on each chunk in parallel across 16
logical processors and merging the results in the end. It was "only" 8x, and not 16x faster as there is
some overhead involved in multithreading.
For fast I/O - let's say you wanted to grab the last 10 records:
$ qsv count --human-readable wcp.csv
2,699,354
$ qsv slice wcp.csv --start -10 | qsv table
Country City AccentCity Region Population Latitude Longitude
zw zibalonkwe Zibalonkwe 06 -19.8333333 27.4666667
zw zibunkululu Zibunkululu 06 -19.6666667 27.6166667
zw ziga Ziga 06 -19.2166667 27.4833333
zw zikamanas village Zikamanas Village 00 -18.2166667 27.95
zw zimbabwe Zimbabwe 07 -20.2666667 30.9166667
zw zimre park Zimre Park 04 -17.8661111 31.2136111
zw ziyakamanas Ziyakamanas 00 -18.2166667 27.95
zw zizalisari Zizalisari 04 -17.7588889 31.0105556
zw zuzumba Zuzumba 06 -20.0333333 27.9333333
zw zvishavane Zvishavane 07 79876 -20.3333333 30.0333333
qsv count took 0.006 seconds and qsv slice, 0.017 seconds! These commands are instantaneous
with an index because for count - the index already precomputed the record count, and with slice,
only the sliced portion has to be parsed - because an index allowed us to jump directly to that
part of the file. It didn't have to scan the entire file to get the last 10 records. For comparison,
without an index, it took 0.25 (41x slower) and 0.66 (39x slower) seconds respectively.
ℹ️ NOTE: Creating/updating an index itself is extremely fast as well. If you want qsv to automatically create and update indices, set the environment var
QSV_AUTOINDEX.
Okay, okay! Let's switch gears and stop obsessing over how fast 🚀 qsv is... let's go back to exploring 🔎 the data set.
Hmmmm... the Population column has a lot of null values. How pervasive is that?
First, let's take a look at 10 "random" rows with sample. We use the --seed parameter
so we get a reproducible random sample. And then, let's display only the Country,
AccentCity and Population columns with the select command.
$ qsv sample --seed 42 10 wcp.csv |
qsv select Country,AccentCity,Population |
qsv table
Country AccentCity Population
ar Colonia Santa Teresa
ro Piscu Scoartei
gr Liáskovo
de Buntenbeck
tr Mehmetçelebi Köyü
pl Trzeciewiec
ar Colonias Unidas
at Koglhof
bg Nadezhda
ru Rabog
Whoops! The sample we got doesn't have population counts. It's quite pervasive. Exactly how many cities have empty (NULL) population counts?
$ qsv frequency wcp.csv --limit 3 | qsv table
field value count
Country ru 176934
Country us 141989
Country cn 117508
City san jose 313
City san antonio 310
City santa rosa 288
AccentCity San Antonio 307
AccentCity Santa Rosa 288
AccentCity Santa Cruz 268
Region 04 143900
Region 02 127736
Region 03 105455
Population (NULL) 2652350
Population 2310 12
Population 2137 11
Latitude 50.8 1128
Latitude 50.95 1076
Latitude 50.6 1043
Longitude 23.1 590
Longitude 23.2 586
Longitude 23.05 575
(The qsv frequency command builds a frequency table for each column in the
CSV data. This one only took 1.8 seconds.)
So it seems that most cities do not have a population count associated with them at all (2,652,350 to be exact). No matter — we can adjust our previous command so that it only shows rows with a population count:
$ qsv search --select Population '[0-9]' wcp.csv |
qsv sample --seed 42 10 |
qsv select Country,AccentCity,Population |
tee sample.csv |
qsv table
Country AccentCity Population
it Isernia 21409
lt Ramygala 1637
ro Band 7599
in Nagapattinam 94247
hn El Negrito 9304
us North Druid Hills 21320
gb Ellesmere Port 67768
bd Parbatipur 48026
sv Apastepeque 5785
ge Lajanurhesi 95
ℹ️ NOTE: The
teecommand reads from standard input and writes to both standard output and one or more files at the same time. We do this so we can create thesample.csvfile we need for the next step, and pipe the same data to theqsv tablecommand.
Why createsample.csv? Even though qsv is blazing-fast, we're just doing an initial investigation and a small 10-row sample is all we need to try out and compose the different CLI commands needed to wrangle the data.
Erk. Which country is sv? What continent? No clue, but datawookie
has a CSV file called country-continent.csv.
$ curl -L https://raw.githubusercontent.com/datawookie/data-diaspora/master/spatial/country-continent-codes.csv > country_continent.csv
$ qsv headers country_continent.csv
1 # https://datahub.io/JohnSnowLabs/country-and-continent-codes-list
Huh!?! That's not what we were expecting. But if you look at the country-continent.csv
file, it starts with a comment with the # character.
$ head -5 country_continent.csv
# https://datahub.io/JohnSnowLabs/country-and-continent-codes-list
continent,code,country,iso2,iso3,number
Asia,AS,"Afghanistan, Islamic Republic of",AF,AFG,4
Europe,EU,"Albania, Republic of",AL,ALB,8
Antarctica,AN,Antarctica (the territory South of 60 deg S),AQ,ATA,10
No worries, qsv got us covered with its QSV_COMMENT_CHAR environment variable. Setting it
to # tells qsv to ignore any lines in the CSV - may it be before the header, or even in the data
part of the CSV, that starts with the character we set it to.
$ export QSV_COMMENT_CHAR='#'
$ qsv headers country_continent.csv
1 continent
2 code
3 country
4 iso2
5 iso3
6 number
That's more like it. We can now do a join to see which countries and continents these are:
$ qsv join --ignore-case Country sample.csv iso2 country_continent.csv | qsv table
Country AccentCity Population continent code country iso2 iso3 number
it Isernia 21409 Europe EU Italy, Italian Republic IT ITA 380
lt Ramygala 1637 Europe EU Lithuania, Republic of LT LTU 440
ro Band 7599 Europe EU Romania RO ROU 642
in Nagapattinam 94247 Asia AS India, Republic of IN IND 356
hn El Negrito 9304 North America NA Honduras, Republic of HN HND 340
us North Druid Hills 21320 North America NA United States of America US USA 840
gb Ellesmere Port 67768 Europe EU United Kingdom of Great Britain & Northern Ireland GB GBR 826
bd Parbatipur 48026 Asia AS Bangladesh, People's Republic of BD BGD 50
sv Apastepeque 5785 North America NA El Salvador, Republic of SV SLV 222
ge Lajanurhesi 95 Europe EU Georgia GE GEO 268
ge Lajanurhesi 95 Asia AS Georgia GE GEO 268
sv is El Salvador - never would have guessed that. Thing is, now we have several unneeded
columns, and the column names case formats are not consistent. Also, there are two records
for Lajanurhesi - for both Europe and Asia. This is because Georgia spans both continents.
We're primarily interested in unique cities per country for the purposes of this tour,
so we need to filter these out.
Also, apart from renaming the columns, I want to reorder them to "City, Population, Country, Continent".
No worries. Let's use the select (so we only get the columns we need, in the order we want),
dedup (so we only get unique County/City combinations) and rename (columns in titlecase) commands:
$ qsv join --ignore-case Country sample.csv iso2 country_continent.csv |
qsv select 'AccentCity,Population,country,continent' |
qsv dedup --select 'country,AccentCity' |
qsv rename City,Population,Country,Continent |
qsv table
City Population Country Continent
Parbatipur 48026 Bangladesh, People's Republic of Asia
Apastepeque 5785 El Salvador, Republic of North America
Lajanurhesi 95 Georgia Asia
El Negrito 9304 Honduras, Republic of North America
Nagapattinam 94247 India, Republic of Asia
Isernia 21409 Italy, Italian Republic Europe
Ramygala 1637 Lithuania, Republic of Europe
Band 7599 Romania Europe
Ellesmere Port 67768 United Kingdom of Great Britain & Northern Ireland Europe
North Druid Hills 21320 United States of America North America
Nice! Notice the data is now sorted by Country,City too! That's because dedup first sorts the
CSV records (by internally calling the qsv sort command) to find duplicates.
Now that we've composed all the commands we need, perhaps we can do this with the original CSV data?
Not the tiny 10-row sample.csv file, but all 2.7 million rows in the 124MB wcp.csv file?!
Indeed we can — because qsv is designed for speed - written in Rust with
amortized memory allocations, using the
performance-focused mimalloc allocator.
$ qsv join --ignore-case Country wcp.csv iso2 country_continent.csv |
qsv search --select Population '[0-9]' |
qsv select 'AccentCity,Population,country,continent,Latitude,Longitude' |
qsv dedup --select 'country,AccentCity,Latitude,Longitude' --dupes-output wcp_dupes.csv |
qsv rename City,Population,Country,Continent,Latitude,Longitude --output wcp_countrycontinent.csv
$ qsv sample 10 --seed 33 wcp_countrycontinent.csv | qsv table
City Population Country Continent Latitude Longitude
Santa Catalina 2727 Philippines, Republic of the Asia 16.0822222 120.6097222
Azacualpa 1258 Honduras, Republic of North America 14.7166667 -88.1
Solana 2984 Philippines, Republic of the Asia 8.6230556 124.7705556
Sungai Besar 26939 Malaysia Asia 3.6666667 100.9833333
Bad Nenndorf 10323 Germany, Federal Republic of Europe 52.3333333 9.3666667
Dalwangan 4906 Philippines, Republic of the Asia 8.2030556 125.0416667
Sharonville 13250 United States of America North America 39.2680556 -84.4133333
El Calvario 557 Colombia, Republic of South America 4.3547222 -73.7091667
Kunoy 70 Faroe Islands Europe 62.2833333 -6.6666667
Lufkin 33667 United States of America North America 31.3380556 -94.7288889
$ qsv count -H wcp_countrycontinent.csv
47,004
$ qsv count -H wcp-dupes.,csv
5,155
We fine-tuned dedup by adding Latitude and Longitude as there may be
multiple cities with the same name in a country. We also specified the
dupes-output option so we can have a separate CSV of the duplicate records
it removed.
We're also just interested in cities with population counts. So we used search
with the regular expression [0-9]. This cuts down the file to 47,004 rows.
The whole thing took ~5 seconds on my machine. The performance of join,
in particular, comes from constructing a SIMD-accelerated hash index of one of the CSV
files. The join command does an inner join by default, but it also has left,
right and full outer, cross, anti and semi join support too. All from the command line,
without having to load the files into a database, index them, to do a SQL join.
Finally, can we create a CSV file for each country of all its cities? Yes we can, with
the partition command (and it took just 0.04 seconds to create all 211 country-city files!):
$ qsv partition Country bycountry wcp_countrycontinent.csv
$ cd bycountry
$ ls -1shS
total 164M
320K UnitedStatesofAmerica.csv
264K PhilippinesRepublicofthe.csv
256K RussianFederation.csv
172K IndiaRepublicof.csv
...
4.0K DjiboutiRepublicof.csv
4.0K Aruba.csv
4.0K Anguilla.csv
4.0K Gibraltar.csv
4.0K Ukraine.csv
Examining the USA csv file:
$ qsv stats --everything UnitedStatesofAmerica.csv | qsv table --output usa-cities-stats.csv
$ less -S usa-cities-stats.csv
field type sum min max min_length max_length mean stddev variance lower_fence q1 q2_median q3 iqr upper_fence skew mode cardinality nullcount
City String Abbeville Zionsville 3 26 Springfield 3439 0
Population Integer 179123400 216 8107916 3 7 42903.80838323359 167752.88891786628 28141031740.28998 -24217.5 12081 19235 36280 24199 72578.5 0.4232798946578281 10576,10945,11971,12115,13219,13250,8771,9944 3981 0
Country String United States of America United States of America 24 24 United States of America 1 0
Continent String North America North America 13 13 North America 1 0
Latitude Float 158455.7901657997 17.9677778 71.2905556 10 10 37.95348267444306 6.0032154906925355 36.03859622769082 22.244444449999992 34.0552778 39.4694444 41.9291667 7.873888900000004 53.740000050000006 -0.7575748669562047 42.0333333 4010 0
Longitude Float -377616.7797696997 -165.4063889 -65.3013889 11 12 -90.44713287897018 17.2089567990395 296.1481941112077 -128.2138889 -97.4863889 -86.0341667 -77.0013889 20.485 -46.273888899999996 -0.769302793394743 -118.3516667,-71.0666667,-71.3972222,-71.4166667,-83.1500000 4074 0
Hhhmmm... clearly the worldcitiespop.csv file from the Data Science Toolkit does not have comprehensive coverage of City populations.
The US population is far more than 179,123,400 (Population sum) and 3,439 cities (City cardinality).
Perhaps we can get population info elsewhere with the fetch command...
But that's another tour by itself! 😄
Footnotes
-
Timings collected by setting
QSV_LOG_LEVEL='debug'on a Ryzen 4800H laptop (8 physical/16 logical cores) running Windows 11 with 32gb of memory and a 1 TB SSD. ↩ ↩2 -
For example, running
qsv statson a CSV export of ALL of NYC's available 311 data from 2010 to Mar 2022 (27.8M rows, 16gb) took just 22.4 seconds with an index (which actually took longer to create - 39 seconds to create a 223mb index), and its memory footprint remained the same, pinning all 16 logical processors near 100% utilization on my Ryzen 7 4800H laptop with 32gb memory and 1 TB SSD. ↩