Add an env_prefix derive option for a consistent prefix for environment variables

[gibfahn]

Maintainer's notes

• Proposed fix is to add a Command::next_env_prefix and Arg::env_prefix and is modeled after help_heading
  • Command::next_env_prefixapplies to all future Args and is set during .arg()
  • An explicit Arg::env_prefix takes precedence over Command::next_env_prefix
  • Derives will need to capture and restore state
  • https://github.com/clap-rs/clap/issues/1807 will allow derives to change the Command::next_env_prefix part-way through struct definition
  • Either we generate the env variables during _build (env will internally need to be a Cow), leaving help generation and env variable look up unchanged or we do it on lookup/display

---

Please complete the following tasks

• [X] I have searched the discussions
• [X] I have searched the existing issues

Clap Version

v3.0.0-rc.8

Describe your use case

This was previously discussed in https://github.com/TeXitoi/structopt/issues/370.

Basically the request is to be able to allow a specific prefix to be appended to all #[clap(env)] derived environment variable names.

Describe the solution you'd like

Before:

#[derive(Parser)]
pub struct Opts {
    #[clap(long, env = "FOO_CONFIG_DIR")]
    pub config_dir: String,
    #[clap(long, env = "FOO_TEMP_DIR")]
    pub temp_dir: String,
    #[clap(long, env = "FOO_CACHE_DIR")]
    pub cache_dir: String,
    #[clap(long)]
    pub isolated: bool
}

After:

#[derive(Parser)]
#[clap(env_prefix)]
pub struct Opts {
    #[clap(long, env)]
    pub config_dir: String,
    #[clap(long, env)]
    pub temp_dir: String,
    #[clap(long, env)]
    pub cache_dir: String,
    #[clap(long)]
    pub isolated: bool
}

(where a custom prefix could be set by #[clap(env_prefix="foo")], but the default was the #[clap(name)])

Alternatives, if applicable

I don't think there's an alternative solution to manually setting the environment variable for every option that needs it.

There was a suggestion in the above issue about allowing this to be calculated via a function vs a string literal, but I can't think of a good use-case for it, so it seems unnecessary for this feature.

Additional Context

Examples from the above issue:

rust-analyzer looks for RA_LOG var, rustc looks for RUSC_LOG var, sccache looks for SCCACHE_ENDPOINT, SCCACHE_CACHE_SIZE, SCCACHE_REDIS and other....

[epage]

Wanted to double check. Did you mean for your "after" to be:

#[derive(Parser)]
#[clap(env_prefix)]
pub struct Opts {
    #[clap(long, env)]
    pub config_dir: String,
    #[clap(long, env)]
    pub temp_dir: String,
    #[clap(long, env)]
    pub cache_dir: String,
    #[clap(long)]
    pub isolated: bool
}

• #[clap(env)] is being specified
• I added --isolated to contrast with the use of env

[epage]

If we do this, we'd need to make clear that env_prefix would not propagate through flatten / subcommand unlike most other attributes that get set. I'm assuming this limitation isn't enough to hold back making it easier to follow common env variable practices.

[gibfahn]

Wanted to double check. Did you mean for your "after" to be:

Whoops, yes my bad, you'd still specify env. I updated the original example to match your correction.

If we do this, we'd need to make clear that env_prefix would not propagate through flatten / subcommand unlike most other attributes that get set.

Why wouldn't we want it to propagate? I can see that making the user be explicit is probably a better idea, but I could also see mycmd mysubccmd --myarg being converted to $MYCMD_MYSUBCMD_MYARG (for example), and I'm not sure what the issue would be with flatten used outside of subcommands.

I'm assuming this limitation isn't enough to hold back making it easier to follow common env variable practices.

Definitely not an issue to specify env_prefix in more places either way, especially if the default values mostly work.

[epage]

If we do this, we'd need to make clear that env_prefix would not propagate through flatten / subcommand unlike most other attributes that get set.

Why wouldn't we want it to propagate? I can see that making the user be explicit is probably a better idea, but I could also see mycmd mysubccmd --myarg being converted to $MYCMD_MYSUBCMD_MYARG (for example), and I'm not sure what the issue would be with flatten used outside of subcommands.

Sorry I didn't go into more details on this. We can't make derives for different structs interact at build-time. We have to do all of this at runtime.

Our options are

• Balloon out the derive API for each new piece of information we want to communicate, either breaking compatibility or having the new functions on the trait call into the old ones with defaults
• Try to figure out which arguments were added during the flatten and modify them accordingly
• Make this a part of the builder API.

In thinking about this, maybe it could be worthwhile to add this to the builder API.

[gibfahn]

In thinking about this, maybe it could be worthwhile to add this to the builder API.

Adding this to the builder API would basically mean that #[clap(env)] would call something like .env(default_env!()) the way #[clap(version)] calls .version(crate_version!()) today?

If so that seems reasonable to me.

Sorry I didn't go into more details on this. We can't make derives for different structs interact at build-time. We have to do all of this at runtime.

For what it's worth I think it would be fine to just have #[clap(env)] work for top-level options, and leave it as a future enhancement to deal with propagating.

[epage]

For what it's worth I think it would be fine to just have #[clap(env)] work for top-level options, and leave it as a future enhancement to deal with propagating.

We want to avoid this. Someone refactoring their app from a single Struct to multiple Structs would then get surprised by it breaking.

[epage]

Alright, so the proposal is we add an App::env_prefix.

Questions

• How should Args override this, e.g. reading from a general purpose env variable
• Should we allow changing the env prefix? Maybe like we do with help_heading where its stateful?
  • Modeling after help_heading would allow overriding the prefix in limited cases
  • For derive users, in https://github.com/clap-rs/clap/issues/1807 we propose a clap::app attribute on fields so App::env_prefix can be called
• Should we apply this during _build or when doing look ups
  • _build is better for sharing logic with help, man pages, etc
  • On lookup is better for performance
  • We could switch to a two-tier build model where get_matches does a minimal build and everything else (like https://github.com/clap-rs/clap/issues/2911) gets a full build.

So in writing up these questions, I think we've got our answers.

[Stargateur]

If we do this, we'd need to make clear that env_prefix would not propagate through flatten / subcommand unlike most other attributes that get set. I'm assuming this limitation isn't enough to hold back making it easier to follow common env variable practices.

It's would be amazing that flatten to do it automatically for example I would like:

#[derive(Debug, Args, Deserialize)]
pub struct PostgreSQL {
  #[clap(long, env)]
  pub host: Option<String>,
}

#[derive(Parser, Debug)]
#[clap(author, version, about)]
#[clap(env_prefix)]
pub struct ConfigArgs {
  #[clap(flatten)]
  pub postgresql: PostgreSQL,
}

This mean host would have the env {NAME}_POSTGRESQL_HOST, if it's not possible it would be great to be able to use env_prefix more than one like:

#[derive(Debug, Args, Deserialize)]
#[clap(env_prefix(name, struct_name))]
pub struct PostgreSQL {
  #[clap(long, env)]
  pub host: Option<String>,
}

[epage]

It's would be amazing that flatten to do it automatically for example I would like:

Note that my later post shifted this from being a derive feature to a builder feature which means it could work across flatten.

[Maher4Ever]

@epage Is this planned to be implemented soon?

[epage]

It has the E-medium tag which means the current design is accepted for moving forward. However, my focus is on other area (iterating on the API for reducing binary size, reducing compile times, and improving flexibility). If someone is willing to step up to implement it, I'd be willing to review and merge the PRs.

[eighty4]

I've been using #[arg(...)] macro yet have seen use of clap interchangeably in other sources. Although it's not on docs.rs, switching out arg for clap works. However, neither arg nor clap permits an env flag (yet I've also seen env used in other sources). Quite confused!

Is there currently a way to permit a required option to be specified with an environment variable or is the environment variable feature included with env_prefix?

Here's an example of my current usage:

https://github.com/eighty4/cquill/blob/main/src/main.rs

Looking for this env_prefix feature led me to this ticket, hoping to have CQUILL_CQL_DIR= override the --cql-dir arg. Is this currently possible with an env flag that would support CQL_DIR=?

[epage]

I've been using #[arg(...)] macro yet have seen use of clap interchangeably in other sources. Although it's not on docs.rs, switching out arg for clap works.

clap is deprecated in favor of the more specialized attributes

However, neither arg nor clap permits an env flag (yet I've also seen env used in other sources). Quite confused!

If you search on docs.rs, you will find Arg::env which is documented as being behind the env feature flag. If you run cargo add clap -F env, it should start to work. As for any derive aspect of this documentation, see the discussion on the topic

[epage]

Alternatively, #5050 would allow doing this without any other new feature