Skip to content

architecture.md

Latest commit

 

History

History
315 lines (233 loc) · 15.6 KB

architecture.md

File metadata and controls

315 lines (233 loc) · 15.6 KB
Raw

Architecture

Contributor-facing architecture reference for cx. Start here before adding a command, debugging the execution pipeline, or reviewing a PR.

Execution flow

Every cx invocation follows a seven-step pipeline:

CLI parsing ──> Config resolution ──> Target building ──> Fan-out
                                                            │
                                              ┌─────────────┘
                                              v
                                        Result merging ──> Output rendering ──> Spilling
Step Owner Key function / type
1. CLI parsing src/main.rs Cli / Commands (Clap derive)
2. Config resolution src/config.rs resolve_all() → Vec<ResolvedConfig>
3. Target building src/execution.rs build_targets() → Vec<Arc<ExecutionTarget>>
4. Fan-out src/execution.rs fan_out(targets, |t| async { ... })
5. Result merging src/execution.rs or src/commands/dataprime/mod.rs merge_tagged_results() / merge_results()
6. Output rendering Command module match output { Text | Json | Toon }
7. Spilling src/spill.rs maybe_spill() (toon mode, DataPrime only)

Step details

1. CLI parsing -- main.rs uses two separate Clap parsers. ProfilesCli handles cx profiles without global API flags (no credentials needed). Cli handles everything else with global --profile, --api-key, --region, and --output flags. Commands that don't need credentials (cleanup, dataprime list/show, schema) are dispatched early before config resolution.

The Cli struct uses a custom help_template with an after_help string to display grouped command categories (Query, Observe, Detect & Respond, etc.) instead of Clap's default flat subcommand listing. This provides a curated help experience without relying on next_help_heading or flatten.

2. Config resolution -- resolve_all() resolves one or more profile names into ResolvedConfig values. Each resolution loads ~/.cx/profiles/<name>.toml, applies CLI/env overrides, and obtains a bearer token (API key from file/keyring, or OAuth with automatic refresh). Precedence: CLI flags > env vars > profile file > global config defaults.

3. Target building -- build_targets() wraps each ResolvedConfig in an ExecutionTarget that owns a pre-built CxClient HTTP client. Targets are Arc-wrapped for cheap cloning into async tasks.

4. Fan-out -- fan_out() runs the same async closure concurrently across all targets using futures::join_all(). Returns Vec<(profile_name, Result<T>)>. Errors are per-profile -- one failing profile doesn't block others.

5-7. Merging, rendering, and spilling differ by command archetype. See below.

Command archetypes

Commands fall into two patterns. New commands should follow one of these.

Archetype A: DataPrime-based

Commands: logs, spans, dataprime query

These delegate to the shared pipeline in commands::dataprime::run_query(), which handles fan-out, merge, and render generically. The command module provides only a source-specific text renderer.

commands::logs::run()
    └─> dataprime::run_query(targets, query, "logs", ..., Some(render_log_text))
            ├─> fan_out(targets, |t| execute_query(t, ...))
            ├─> merge_results(per_profile, include_profile)
            └─> render_results(merged, output, max_direct, temp_dir, text_renderer)

Adding a new DataPrime command requires:

  1. A command directory (e.g., src/commands/logs/) with a mod.rs containing a text renderer function matching fn(&MergedResults) -> Result<()>
  2. A run() function that calls dataprime::run_query() with the source name and renderer
  3. CLI definition in main.rs and dispatch in the match cli.command block

Reference: src/commands/logs/mod.rs -- the entire module is ~130 lines, most of which is the text renderer. The run() function is a single delegation call.

Archetype B: REST-based

Commands: alerts, cases, dashboards, metrics, search-fields, notifications, webhooks, parsing-rules, enrichments, integrations, iam, usage, tco, retentions, archive, slos, views, e2m, recording-rules

These manage their own fan-out, merge, and render inline. Each subcommand function follows the same shape:

pub async fn run_list(targets, ..., output) -> Result<()> {
    let include_profile = targets.len() > 1;

    // Fan-out: call the API across all profiles
    let per_profile = fan_out(targets, |t| async move {
        let api = AlertsApi::new(&t.client);
        Ok(api.list().await?)
    }).await;

    // Merge: collect results, print errors to stderr
    let mut all_items = Vec::new();
    for (profile, result) in per_profile {
        match result {
            Ok(resp) => { /* filter, push to all_items */ }
            Err(e) => eprintln!("error from profile '{profile}': {e:#}"),
        }
    }

    // Render: match on output format
    match output {
        OutputFormat::Json => { /* serde_json::to_string_pretty */ }
        OutputFormat::Toon => { /* toon_encode */ }
        OutputFormat::Text => { /* Table::new(rows) */ }
    }
}

Adding a new REST command requires:

  1. A command directory (src/commands/<resource>/) with:
    • api.rs -- a <Resource>Api struct wrapping &CxClient
    • mod.rs -- run_<subcommand>() functions, with pub mod api; at the top
  2. CLI definition in main.rs and dispatch

Reference: src/commands/dashboards/ -- clean example using render::* helpers for all three output formats.

Wrapper enum pattern

Several CLI commands group multiple related domains under a single top-level command. These use a wrapper enum in main.rs that nests further subcommand enums:

// Top-level wrapper in Commands enum
Iam {
    #[command(subcommand)]
    cmd: IamCmd,
},

// Wrapper enum groups related sub-domains
#[derive(Subcommand)]
enum IamCmd {
    ApiKeys { #[command(subcommand)] cmd: ApiKeysCmd },
    Roles { #[command(subcommand)] cmd: RolesCmd },
    Scopes { #[command(subcommand)] cmd: ScopesCmd },
    Users { #[command(subcommand)] cmd: UsersCmd },
    Groups { #[command(subcommand)] cmd: GroupsCmd },
    IpAccess { #[command(subcommand)] cmd: IpAccessCmd },
}

This pattern is used by: notifications (connectors, routers, presets, test), webhooks (list/get/types + actions), enrichments (rules + custom), integrations (list/get + extensions + contextual-data), and iam (api-keys, roles, scopes, users, groups, ip-access).

Each sub-domain still has its own directory (src/commands/<sub_domain>/) with mod.rs (handler) and api.rs (HTTP client). The wrapper only affects CLI wiring and dispatch in main.rs.

Help display

The Cli struct uses a custom help_template with an after_help string to display grouped command categories rather than Clap's default flat subcommand listing. The after_help text is a manually maintained string that mirrors the Commands enum. When adding a new command, update the after_help string to place it in the correct category group.

Destructive operation confirmation

All write operations (create, update, delete, enable, disable, set) require interactive confirmation. The confirmation logic lives in src/safety.rs:

  1. If --yes is passed, the operation proceeds immediately
  2. If stdin is not a terminal, the operation fails with a message directing the user to pass --yes
  3. Otherwise, inquire::Confirm prompts the user (default: No)

The --yes flag is global on the Cli struct and available as yes in all match arms. Each destructive subcommand is tagged [requires --yes] in its doc comment, which surfaces in --help output.

When adding new destructive operations, call confirm_destructive(action, yes, agent_mode) before the handler invocation. See adding-a-command.md for the full pattern.

Output rendering

All commands support three output formats controlled by --output / OutputFormat:

Text

Human-readable output using tabled for tabular data and colored for severity/status highlighting. Multi-profile queries add a "Profile" column. REST commands use render::render_table() which dynamically builds tables via tabled::builder::Builder -- the Profile column is conditionally included based on include_profile, so no duplicate struct definitions are needed.

DataPrime commands use custom text renderers (e.g., render_log_text) that print <timestamp> [<severity>] <message> per row.

JSON

Raw API responses pretty-printed via serde_json::to_string_pretty. No transformation applied.

TOON

Token-optimized format for AI consumers:

  1. TOON encoding -- all toon output uses toon_format::encode_default() for compact serialization
  2. Metadata stripping (DataPrime only) -- transform_for_toon() renames keys (metadata -> $m, labels -> $l, userData -> $d) and removes noisy metadata fields (branchid, priorityclass, *TimestampMicros, etc.)
  3. Spilling (DataPrime non-aggregates only) -- maybe_spill() checks serialized size against max_dataprime_direct_output_size (default 100 KiB). If exceeded, writes to cx_results_<hash>.json in temp_dir and prints the path instead.

Multi-profile pattern

A single boolean controls profile-aware behavior throughout the codebase:

let include_profile = targets.len() > 1;

When true:

  • Merging: tag_rows() / merge_tagged_results() inject "profile": "<name>" into each JSON row
  • Text output: commands use the multi-profile row struct (with Profile column)
  • Errors: printed to stderr with "error from profile '<name>'" prefix

When false:

  • No profile field is injected
  • Single-profile row struct is used
  • Errors still print to stderr

The generic version lives in execution.rs (tag_rows, merge_tagged_results). DataPrime commands use a specialized dataprime::merge_results() that also handles is_aggregate detection and warning collection.

Error handling

Two error systems coexist, each serving a different layer:

API layer: CxError

Defined in src/error.rs. A typed enum used throughout command api.rs modules:

Variant When Example
Auth(String) 401/403 responses Invalid API key
Api { status, message } Other non-2xx responses 404 alert not found
Http(reqwest::Error) Network failures Connection refused
Json(serde_json::Error) Deserialization failures Unexpected response shape
Io(std::io::Error) File I/O failures Config file unreadable

CxError enables pattern matching on specific API errors:

match api.get(&id).await {
    Ok(val) => Ok(val),
    Err(CxError::Api { status: 404, .. }) => Ok(api.get_by_version_id(&id).await?),
    Err(e) => Err(anyhow::Error::from(e)),
}

Command layer: anyhow::Result

All command functions return anyhow::Result. CxError converts automatically via the ? operator. Use .context() to add human-readable messages:

serde_yaml::from_str(yaml).context("Failed to parse embedded dataprime documentation")?;

Fan-out errors

Errors during fan-out are per-profile and non-fatal:

  • Printed to stderr: eprintln!("error from profile '{profile}': {e:#}")
  • Successful profiles continue normally
  • The command exits 0 if at least one profile succeeds

The API client

CxClient (src/api_client.rs) is a thin reqwest wrapper pre-configured with Bearer auth:

Method Use case
post_raw(path, body) NDJSON/streaming endpoints (DataPrime) -- returns raw text
post<T>(path, body) Typed JSON endpoints -- deserializes response
get<T>(path, params) GET with query params
post_empty<T>(path, params) State-change endpoints (enable/disable) -- handles empty 200/204

All methods run through checked_text() which maps HTTP status codes to CxError variants.

API modules (src/commands/<resource>/api.rs) wrap CxClient in domain-specific structs:

pub struct AlertsApi<'a> {
    client: &'a CxClient,
}

impl<'a> AlertsApi<'a> {
    pub fn new(client: &'a CxClient) -> Self { Self { client } }
    pub async fn list(&self) -> crate::error::Result<ListResponse> { ... }
}

Naming conventions

Element Convention Example
Source files snake_case.rs search_fields.rs
Command functions run_<subcommand>() run_list(), run_get(), run_query()
API structs <Resource>Api AlertsApi, MetricsApi
Render helpers render::render_table, render::render_json, etc. Shared text/JSON output
Command directory src/commands/<resource>/ mod.rs for the handler, api.rs for REST commands
Error types CxError for API, anyhow::Result for commands --

Module map

Each CLI command owns a directory under src/commands/. REST commands keep their HTTP client alongside the handler in api.rs. Cross-cutting infrastructure stays at the top of src/. The shared HTTP base lives at src/api_client.rs.

src/
├── main.rs              # CLI definition (Clap) + dispatch + help_template/after_help
├── lib.rs               # Module re-exports
├── api_client.rs        # CxClient HTTP wrapper (Bearer auth, REST + NDJSON)
├── safety.rs            # Read-only enforcement, agent-mode detection, destructive op confirmation
├── config.rs            # Config/profile loading, resolution, Region enum
├── execution.rs         # ExecutionTarget, fan_out(), tag_rows(), merge_tagged_results()
├── render.rs            # Shared rendering helpers (render_table, render_json, bool_display, etc.)
├── error.rs             # CxError enum
├── spill.rs             # TOON output spilling + transform_for_toon()
├── time.rs              # Relative/absolute timestamp parsing
├── tier.rs              # Tier enum (FrequentSearch | Archive)
├── oauth.rs             # OAuth 2.0 + OIDC browser login flow
├── keyring_store.rs     # OS keyring read/write
├── serde_helpers.rs     # Shared serde deserializers (string_or_number, etc.)
└── commands/
    ├── dataprime/       # Shared DataPrime pipeline + docs subcommands
    │   ├── mod.rs       #   handler + docs (list/show)
    │   ├── api.rs       #   DataPrime query API (NDJSON streaming)
    │   └── semantic_search.rs  # Semantic search API (used by metrics + search-fields)
    ├── logs/mod.rs      # Log query (DataPrime archetype)
    ├── spans/mod.rs     # Span query (DataPrime archetype)
    ├── metrics/         # PromQL commands (REST archetype)
    │   ├── mod.rs
    │   └── api.rs       #   PromQL query APIs
    ├── alerts/          # Alert management (REST archetype)
    │   ├── mod.rs
    │   └── api.rs       #   Alerts CRUD API
    ├── dashboards/      # Dashboard commands (REST archetype)
    │   ├── mod.rs
    │   └── api.rs       #   Dashboards API
    ├── search_fields/mod.rs  # Semantic field lookup (uses dataprime::semantic_search)
    ├── schema/mod.rs    # cx schema - JSON command tree for agent discovery
    ├── profiles/mod.rs  # Profile management (no API calls)
    ├── cleanup/mod.rs   # Temp file cleanup (no API calls)
    └── ...              # One module per REST domain (cases, notifications,
                         #   webhooks, rules, enrichments, integrations, iam,
                         #   usage, tco, retentions, archive, slos,
                         #   views, e2m, recording_rules)