| 1 | ---
|
| 2 | default_highlighter: oils-sh
|
| 3 | ---
|
| 4 |
|
| 5 | YSH Design: Streams, Tables and Processes (and Trees?)
|
| 6 | ======================================================
|
| 7 |
|
| 8 | *(July 2024)*
|
| 9 |
|
| 10 | This is a long, "unified/orthogonal" design for:
|
| 11 |
|
| 12 | - Streams: [awk]($xref) delimited lines, regexes
|
| 13 | - Tables: like data frames with R's dplyr or Pandas, but with the "exterior"
|
| 14 | TSV8 format
|
| 15 | - Processes: xargs -P in parallel
|
| 16 |
|
| 17 | There's also a relation to:
|
| 18 |
|
| 19 | - Trees: `jq`, which will be covered elsewhere.
|
| 20 |
|
| 21 | It's a layered design. That means we need some underlying mechanisms:
|
| 22 |
|
| 23 | - `eval` and positional args `$1 $2 $3`
|
| 24 | - `ctx` builtin
|
| 25 | - Data langauges: TSV8
|
| 26 | - Process pool / event loop primitive
|
| 27 |
|
| 28 | It will link to:
|
| 29 |
|
| 30 | - Oils blog posts (old)
|
| 31 | - Zulip threads (recent)
|
| 32 | - Other related projects (many of them)
|
| 33 |
|
| 34 | <div id="toc">
|
| 35 | </div>
|
| 36 |
|
| 37 | ## Background / References
|
| 38 |
|
| 39 | - Shell, Awk, and Make Should be Combined (2016)
|
| 40 | - this is the Awk part!
|
| 41 |
|
| 42 | - What is a Data Frame? (2018)
|
| 43 |
|
| 44 | - Sketches of YSH Features (June 2023) - can we express things in YSH?
|
| 45 | - Zulip: Oils Layering / Self-hosting
|
| 46 |
|
| 47 | - Language Compositionality Test: J8 Lines
|
| 48 | - This whole thing is a compositionality test
|
| 49 |
|
| 50 | - read --split
|
| 51 | - more feedback from Aidan and Samuel
|
| 52 |
|
| 53 | - What is a Data Frame?
|
| 54 |
|
| 55 | - jq in jq thread
|
| 56 |
|
| 57 | Old wiki pages:
|
| 58 |
|
| 59 | - <https://github.com/oilshell/oil/wiki/Structured-Data-in-Oil>
|
| 60 | - uxy - closest I think - <https://github.com/sustrik/uxy>
|
| 61 | - relies on to-json and jq for querying
|
| 62 | - miller - I don't like their language - https://github.com/johnkerl/miller -
|
| 63 | - jc - <https://github.com/kellyjonbrazil/jc>
|
| 64 | - nushell
|
| 65 | - extremely old thing -
|
| 66 |
|
| 67 | We're doing **all of these**.
|
| 68 |
|
| 69 | ## Concrete Use Cases
|
| 70 |
|
| 71 | - benchmarks/* with dplyr
|
| 72 | - wedge report
|
| 73 | - oilshell.org analytics job uses dplyr and ggplot2
|
| 74 |
|
| 75 | ## Intro
|
| 76 |
|
| 77 | ### How much code is it?
|
| 78 |
|
| 79 | - I think this is ~1000 lines of Python and ~1000 lines of YSH (not including tests)
|
| 80 | - It should be small
|
| 81 |
|
| 82 | ### Thanks
|
| 83 |
|
| 84 | - Samuel - two big hints
|
| 85 | - do it in YSH
|
| 86 | - `table` with the `ctx` builtin
|
| 87 | - Aidan
|
| 88 | - `read --split` feedback
|
| 89 |
|
| 90 |
|
| 91 | ## Tools
|
| 92 |
|
| 93 | - awk
|
| 94 | - streams of records - row-wise
|
| 95 | - R
|
| 96 | - column-wise operations on tables
|
| 97 | - `find . -printf '%s %P\n'` - size and path
|
| 98 | - generate text that looks like a table
|
| 99 | - xargs
|
| 100 | - operate on tabular text -- it has a bespoke splitting algorithm
|
| 101 | - Opinionated guide to xargs
|
| 102 | - table in, table out
|
| 103 | - jq - "awk for JSON"
|
| 104 |
|
| 105 |
|
| 106 | ## Concepts
|
| 107 |
|
| 108 | - TSV8
|
| 109 | - aligned format SSV8
|
| 110 | - columns have types, and attributes
|
| 111 | - Lines
|
| 112 | - raw lines like shell
|
| 113 | - J8 lines (which can represent any filename, any unicode or byte string)
|
| 114 | - Tables - can be thought of as:
|
| 115 | - Streams of Rows - shape `[{bytes: 123, path: "foo"}, {}, ...]`
|
| 116 | - this is actually <https://jsonlines.org> , and it fits well with `jq`
|
| 117 | - Columns - shape `{bytes: [], path: []}
|
| 118 |
|
| 119 | ## Underlying Mechanisms in Oils / Primitives
|
| 120 |
|
| 121 | - blocks `value.Block` - `^()` and `{ }`
|
| 122 | - expressions `value.Expr` - `^[]` and 'compute [] where []'
|
| 123 |
|
| 124 | - eval (b, vars={}, positional=[])
|
| 125 |
|
| 126 | - Buffered for loop
|
| 127 | - YSH is now roughly as fast as Awk!
|
| 128 | - `for x in <>`
|
| 129 |
|
| 130 | - "magic awk loop"
|
| 131 |
|
| 132 | with chop {
|
| 133 | for <README.md *.py> {
|
| 134 | echo _line_num _line _filename $1 $2
|
| 135 | }
|
| 136 | }
|
| 137 |
|
| 138 | - positional args $1 $2 $3
|
| 139 | - currently mean "argv stack"
|
| 140 | - or "the captures"
|
| 141 | - this can probably be generalized
|
| 142 |
|
| 143 | - `ctx` builtin
|
| 144 | - `value.Place`
|
| 145 |
|
| 146 | TODO:
|
| 147 |
|
| 148 | - split() like Python, not like shell IFS algorithm
|
| 149 |
|
| 150 | - string formatting ${bytes %.2f}
|
| 151 | - ${bytes %.2f M} Megabytes
|
| 152 | - ${bytes %.2f Mi} Mebibytes
|
| 153 |
|
| 154 | - ${timestamp +'%Y-m-%d'} and strfitime
|
| 155 |
|
| 156 | - this is for
|
| 157 |
|
| 158 | - floating point %e %f %g and printf and strftime
|
| 159 |
|
| 160 | ### Process Pool or Event Loop Primitive?
|
| 161 |
|
| 162 | - if you want to display progress, then you might need an event loop
|
| 163 | - test framework might display progress
|
| 164 |
|
| 165 | ## Matrices - Orthogonal design in these dimensions
|
| 166 |
|
| 167 | - input: lines vs. rows
|
| 168 | - output: string (Str, Template) vs. row vs. block execution (also a row)
|
| 169 | - execution: serial vs. parallel
|
| 170 | - representation: interior vs. exterior !!!
|
| 171 | - Dicts and Lists are interior, but TSV8 is exterior
|
| 172 | - and we have row-wise format, and column-wise format -- this always bugged me
|
| 173 | - exterior: human vs. machine readable
|
| 174 | - TSV8 is both human and machine-readable
|
| 175 | - "aligned" #.ssv8 format is also
|
| 176 | - they are one format named TSV8, with different file extensions. This is
|
| 177 | because it doesn't make too much sense to implement SSV8 without TSV8. The
|
| 178 | latter becomes trivial. So we call the whole thing TSV8.
|
| 179 |
|
| 180 | This means we consider all these conversions
|
| 181 |
|
| 182 | - Line -> Line
|
| 183 | - Line -> Row
|
| 184 | - Row -> Line
|
| 185 | - Row -> Row
|
| 186 |
|
| 187 | ## Concrete Decisions - Matrix cut off
|
| 188 |
|
| 189 | Design might seem very general, but we did make some hard choices.
|
| 190 |
|
| 191 | - push vs. pull
|
| 192 | - everything is "push" style I think
|
| 193 | - buffered vs. unbuffered, everything
|
| 194 |
|
| 195 | - List vs iterators
|
| 196 | - everything is either iterable pipelines, or a List
|
| 197 |
|
| 198 |
|
| 199 | [OSH]: $xref
|
| 200 | [YSH]: $xref
|
| 201 |
|
| 202 |
|
| 203 | ## String World
|
| 204 |
|
| 205 | **THESE ARE ALL THE SAME ALGORITHM**. They just have different names.
|
| 206 |
|
| 207 | - each-line
|
| 208 | - each-row
|
| 209 | - split-by (/d+/, cols=:|a b c|)
|
| 210 | - chop
|
| 211 | - if-match
|
| 212 | - must-match
|
| 213 | - todo
|
| 214 |
|
| 215 | should we also have: if-split-by ? In case there aren't enough columns?
|
| 216 |
|
| 217 | They all take:
|
| 218 |
|
| 219 | - string arg ' '
|
| 220 | - template arg (^"") - `value.Expr`
|
| 221 | - block arg
|
| 222 |
|
| 223 | for the block arg, this applies:
|
| 224 |
|
| 225 | -j 4
|
| 226 | --max-jobs 4
|
| 227 |
|
| 228 | --max-jobs $(cached-nproc)
|
| 229 | --max-jobs $[_nproc - 1]
|
| 230 |
|
| 231 | ### Awk Issues
|
| 232 |
|
| 233 | So we have this
|
| 234 |
|
| 235 | echo begin
|
| 236 | var d = {}
|
| 237 | cat -- @files | split-by (ifs=IFS) {
|
| 238 | echo $2 $1
|
| 239 | call d->accum($1, $2)
|
| 240 | }
|
| 241 | echo end
|
| 242 |
|
| 243 | But then how do we have conditionals:
|
| 244 |
|
| 245 | Filter foo { # does this define a proc? Or a data structure
|
| 246 |
|
| 247 | split-by (ifs=IFS) # is this possible? We register the proc itself?
|
| 248 |
|
| 249 | config split-by (ifs=IFS) # register it
|
| 250 |
|
| 251 | BEGIN {
|
| 252 | var d = {}
|
| 253 | }
|
| 254 | END {
|
| 255 | echo d.sum
|
| 256 | }
|
| 257 |
|
| 258 | when [$1 ~ /d+/] {
|
| 259 | setvar d.sum += $1
|
| 260 | }
|
| 261 |
|
| 262 | }
|
| 263 |
|
| 264 | ## Table World
|
| 265 |
|
| 266 | ### `table` to construct
|
| 267 |
|
| 268 | Actions:
|
| 269 |
|
| 270 | table cat
|
| 271 | table align / table tabify
|
| 272 | table header (cols)
|
| 273 | table slice (1, -1) or (-1, -2) etc.
|
| 274 |
|
| 275 | Subcommands
|
| 276 |
|
| 277 | cols
|
| 278 | types
|
| 279 | attr units
|
| 280 |
|
| 281 | Partial Parsing / Lazy Parsing - TSV8 is designed for this
|
| 282 |
|
| 283 | # we only decode the columns that are necessary
|
| 284 | cat myfile.tsv8 | table --by-col (&out, cols = :|bytes path|)
|
| 285 |
|
| 286 | ## Will writing it in YSH be slow?
|
| 287 |
|
| 288 | - We concentrate on semantics first
|
| 289 | - We can rewrite in Python
|
| 290 | - Better: users can use **exterior** tools with the same interface
|
| 291 | - in some cases
|
| 292 | - they can write an efficient `sort-tsv8` or `join-tsv8` with novel algorithms
|
| 293 | - Most data will be small at first
|
| 294 |
|
| 295 |
|
| 296 | ## Applications
|
| 297 |
|
| 298 | - Shell is shared nothing
|
| 299 | - Scaling to infinity on the biggest clouds
|
| 300 |
|
| 301 |
|
| 302 | ## Extra: Tree World?
|
| 303 |
|
| 304 | This is sort of "expanding the scope" of the project, when we want to reduce scope.
|
| 305 |
|
| 306 | But YSH has both tree-shaped JSON, and table-shaped TSV8, and jq is a nice **bridge** between them.
|
| 307 |
|
| 308 | Streams of Trees (jq)
|
| 309 |
|
| 310 | empty
|
| 311 | this
|
| 312 | this[]
|
| 313 | =>
|
| 314 | select()
|
| 315 | a & b # more than one
|
| 316 |
|
| 317 |
|
| 318 | ## Pie in the Sky
|
| 319 |
|
| 320 | Four types of Data Languages:
|
| 321 |
|
| 322 | - flat strings
|
| 323 | - JSON8 - tree
|
| 324 | - TSV8 - table
|
| 325 | - NIL8 - Lisp Tree
|
| 326 | - HTML/XML - doc tree -- attributed text (similar to Emacs data model)
|
| 327 | - 8ml
|
| 328 |
|
| 329 | Four types of query languaegs:
|
| 330 |
|
| 331 | - regex
|
| 332 | - jq / jshape
|
| 333 | - tsv8
|
| 334 |
|
| 335 |
|
| 336 | ## Appendix
|
| 337 |
|
| 338 | ### Notes on Naming
|
| 339 |
|
| 340 | Considering columns and then rows:
|
| 341 |
|
| 342 | - SQL is "select ... where"
|
| 343 | - dplyr is "select ... filter"
|
| 344 | - YSH is "pick ... where"
|
| 345 | - select is a legacy shell keyword, and pick is shorter
|
| 346 | - or it could be elect in OSH, elect/select in YSH
|
| 347 | - OSH wouldn't support mutate [average = bytes/total] anyway
|
| 348 |
|
| 349 | dplyr:
|
| 350 |
|
| 351 | - summarise vs. summarize vs. summary
|
| 352 |
|
| 353 |
|
| 354 |
|
| 355 |
|
| 356 |
|