Improves startup time when using std-lib (#13842)
Updated summary for commit
[612e0e2](https://github.com/nushell/nushell/pull/13842/commits/612e0e21602f55092bc121bfd07f7c3bf5119e4f)
- While folks are welcome to read through the entire comments, the core
information is summarized here.
# Description
This PR drastically improves startup times of Nushell by only parsing a
single submodule of the Standard Library that provides the `banner` and
`pwd` commands. All other Standard Library commands and submodules are
parsed when imported by the user. This cuts startup times by more than
60%.
At the moment, we have stopped adding to `std-lib` because every
addition adds a small amount to the Nushell startup time.
With this change, we should once again be able to allow new
functionality to be added to the Standard Library without it impacting
`nu` startup times.
# User-Facing Changes
* Nushell now starts about 60% faster
* Breaking change: The `dirs` (Shells) aliases will return a warning
message that it will not be auto-loaded in the following release, along
with instructions on how to restore it (and disable the message)
* The `use std <submodule> *` syntax is available for convenience, but
should be avoided in scripts as it parses the entire `std` module and
all other submodules and places it in scope. The correct syntax to
*just* load a submodule is `use std/<submodule> *` (asterisk optional).
The slash is important. This will be documented.
* `use std *` can be used for convenience to load all of the library but
still incurs the full loading-time.
* `std/dirs`: Semi-breaking change. The `dirs` command replaces the
`show` command. This is more in line with the directory-stack
functionality found in other shells. Existing users will not be impacted
by this as the alias (`shells`) remains the same.
* Breaking-change: Technically a breaking change, but probably only
impacts maintainers of `std`. The virtual path for the standard library
has changed. It could previously be imported using its virtual path (and
technically, this would have been the correct way to do it):
```nu
use NU_STDLIB_VIRTUAL_DIR/std
```
The path is now simply `std/`:
```nu
use std
```
All submodules have moved accordingly.
# Timings
Comparisons below were made:
* In a temporary, clean config directory using `$env.XDG_CONFIG_HOME =
(mktemp -d)`.
* `nu` was run with a release build
* `nu` was run one time to generate the default `config.nu` (etc.) files
- Otherwise timings would include the user-prompt
* The shell was exited and then restarted several times to get timing
samples
(Note: Old timings based on 0.97 rather than 0.98, but in the range of
being accurate)
| Scenario | `$nu.startup-time` |
| --- | --- |
| 0.97.2
([aaaab8e](https://github.com/nushell/nushell/commit/aaaab8e070c644a87bbd7682099e3fe9e6a4b42a))
Without this PR | 23ms - 24ms |
| This PR with deprecated commands | 9ms - <11ms |
| This PR after deprecated commands are removed in following release |
8ms - <10ms |
| Final PR (remove deprecated), using `--no-std-lib` | 6.1ms to 6.4ms |
| Final PR (remove deprecated), using `--no-config-file` | 3.1ms - 3.6ms
|
| Final PR (remove deprecated), using `--no-config-file --no-std-lib` |
1ms - 1.5ms |
*These last two timings point to the opportunity for further
optimization (see comment in thread below (will link once I write it).*
# Implementation details for future maintenance
* `use std banner` is a ridiculously deceptive call. That call parses
and imports *all* of `std` into scope. Simply replacing it with `use
std/core *` is essentially what saves ~14-15ms. This *only* imports the
submodule with the `banner` and `pwd` commands.
* From the code-comments, the reason that `NU_STDLIB_VIRTUAL_DIR` was
used as a prefix was so that there wouldn't be an issue if a user had a
`./std/mod.nu` in the current directory. This does **not** appear to be
an issue. After removing the prefix, I tested with both a relative
module as well as one in the `$env.NU_LIB_DIRS` path, and in all cases
the *internal* `std` still took precedence.
* By removing the prefix, users can now `use std` (and variants) without
requiring that it already be parsed and in scope.
* In the next release, we'll stop autoloading the `dirs` (shells)
functionality. While this only costs an additional 1-1.5ms, I think it's
better moved to the `config.nu` where the user can optionally remove it.
The main reason is its use of aliases (which have also caused issues) -
The `n`, `p`, and `g` short-commands are valuable real-estate, and users
may want to map these to something else.
For this release, there's an `deprecated_dirs` module that is still
autoloaded. As with the top-level commands, use of these will give a
deprecation warning with instructions on how to handle going forward.
To help with this, moved the aliases to their own submodule inside the
`dirs` module.
* Also sneaks in a small change where the top-level `dirs` command is
now the replacement for `dirs show`
* Fixed a double-import of `assert` in `dirs.nu`
* The `show_banner` step is replaced with simply `banner` rather than
re-importing it.
* A `virtual_path` may now be referenced with either a forward-slash or
a backward-slash on Windows. This allows `use std/<submodule>` to work
on all platforms.
# Performance side-notes:
* Future parsing and/or IR improvements should improve performance even
further.
* While the existing load time penalty of `std-lib` was not noticeable
on many systems, Nushell runs on a wide-variety of hardware and OS
platforms. Slower platforms will naturally see a bigger jump in
performance here. For users starting multiple Nushell sessions
frequently (e.g., `tmux`, Zellij, `screen`, et. al.) it is recommended
to keep total startup time (including user configuration) under ~250ms.
# Tests + Formatting
* All tests are green
* Updated tests:
- Removed the test that confirmed that `std` was loaded (since we
don't).
- Removed the `shells` test since it is not autoloaded. Main `dirs.nu`
functionality is tested through `stdlib-test`.
- Many tests assumed that the library was fully loaded, because it was
(even though we didn't intend for it to be). Fixed those tests.
- Tests now import only the necessary submodules (e.g., `use
std/assert`, rather than `use std assert`)
- Some tests *thought* they were loading `std/log`, but were doing so
improperly. This was masked by the now-fixed "load-everything-into-scope
bug". Local CI would pass due the `$env.NU_LOG_<...>` variables being
inherited from the calling process, but would fail in the "clean" GitHub
CI environment. These tests have also been fixed.
* Added additional tests for the changes
# After Submitting
Will update the Standard Library doc page
2024-10-03 13:28:22 +02:00
|
|
|
use std/log
|
|
|
|
use std/assert
|
2023-04-10 22:42:11 +02:00
|
|
|
|
2023-07-02 10:41:33 +02:00
|
|
|
#[before-each]
|
2023-06-10 20:16:17 +02:00
|
|
|
def before-each [] {
|
2023-04-10 22:42:11 +02:00
|
|
|
log debug "Setup is running"
|
|
|
|
{msg: "This is the context"}
|
|
|
|
}
|
|
|
|
|
2023-07-02 10:41:33 +02:00
|
|
|
#[after-each]
|
2023-06-10 20:16:17 +02:00
|
|
|
def after-each [] {
|
2023-04-10 22:42:11 +02:00
|
|
|
log debug $"Teardown is running. Context: ($in)"
|
|
|
|
}
|
|
|
|
|
2023-07-02 10:41:33 +02:00
|
|
|
#[test]
|
|
|
|
def assert_pass [] {
|
2023-04-10 22:42:11 +02:00
|
|
|
log debug $"Assert is running. Context: ($in)"
|
|
|
|
}
|
|
|
|
|
2023-07-02 10:41:33 +02:00
|
|
|
#[ignore]
|
|
|
|
def assert_skip [] {
|
2023-04-10 22:42:11 +02:00
|
|
|
log debug $"Assert is running. Context: ($in)"
|
|
|
|
}
|
|
|
|
|
2023-07-02 10:41:33 +02:00
|
|
|
#[ignore]
|
|
|
|
def assert_fail_skipped_by_default [] {
|
|
|
|
# Change test-skip to test if you want to see what happens if a test fails
|
2023-04-10 22:42:11 +02:00
|
|
|
log debug $"Assert is running. Context: ($in)"
|
|
|
|
assert false
|
|
|
|
}
|
|
|
|
|
2023-06-10 20:16:17 +02:00
|
|
|
def unrelated [] {
|
2023-04-10 22:42:11 +02:00
|
|
|
log error "This should not run"
|
|
|
|
}
|