faucet_source_bigquery/

stream.rs

//! BigQuery query source.
//!
//! Submits the configured SQL statement via `jobs.query` and pages through
//! the result set via `jobs.getQueryResults`. The first response may carry
//! `jobComplete=false` (statement still running on the server side); the
//! source polls `getQueryResults` until BigQuery flips that flag, exactly
//! mirroring the behaviour of `gcp_bigquery_client::Client::job().query_all`
//! without giving up the row-level access we need for incremental
//! [`StreamPage`]s.

use crate::config::BigQuerySourceConfig;
use crate::convert::row_to_json;
use async_trait::async_trait;
use faucet_common_bigquery::build_client;
use faucet_core::util::substitute_context_bind_params;
use faucet_core::{DatasetDescriptor, FaucetError, Stream, StreamPage};
use gcp_bigquery_client::Client;
use gcp_bigquery_client::dataset::ListOptions as DatasetListOptions;
use gcp_bigquery_client::model::field_type::FieldType;
use gcp_bigquery_client::model::get_query_results_parameters::GetQueryResultsParameters;
use gcp_bigquery_client::model::query_parameter::QueryParameter;
use gcp_bigquery_client::model::query_parameter_type::QueryParameterType;
use gcp_bigquery_client::model::query_parameter_value::QueryParameterValue;
use gcp_bigquery_client::model::query_request::QueryRequest;
use gcp_bigquery_client::model::query_response::QueryResponse;
use gcp_bigquery_client::model::table_field_schema::TableFieldSchema;
use gcp_bigquery_client::model::table_row::TableRow;
use gcp_bigquery_client::table::ListOptions as TableListOptions;
use serde_json::{Value, json};
use std::collections::HashMap;
use std::pin::Pin;
use std::time::Duration;

/// Hard cap on the total number of tables [`Source::discover`] enumerates
/// across all datasets in the project. Discovery is a preflight convenience —
/// a project with thousands of tables should not turn it into an API storm.
/// When the cap is hit, enumeration stops and a warning names the cap.
const MAX_DISCOVER_TABLES: usize = 500;

/// Hard cap on per-table `tables.get` schema/row-count fetches during
/// [`Source::discover`]. Tables beyond this cap are still emitted (name +
/// `config_patch`), just without a `schema` / `estimated_rows`, and a warning
/// names the cap.
const MAX_DISCOVER_SCHEMA_FETCHES: usize = 100;

/// A source that runs a SQL query against BigQuery and yields rows as JSON.
pub struct BigQuerySource {
    config: BigQuerySourceConfig,
    client: Client,
}

impl BigQuerySource {
    /// Create a new BigQuery source from the given configuration.
    ///
    /// Initialises the underlying BigQuery client and exchanges credentials
    /// for an OAuth token. Returns [`FaucetError::Auth`] on credential
    /// failures.
    pub async fn new(config: BigQuerySourceConfig) -> Result<Self, FaucetError> {
        faucet_core::validate_batch_size(config.batch_size)?;
        let client = build_client(&config.auth).await?;
        Ok(Self { config, client })
    }

    /// Construct a source from a pre-built BigQuery client.
    ///
    /// Low-level escape hatch for callers that build their own
    /// [`gcp_bigquery_client::Client`] — for example to target the
    /// [`bigquery-emulator`](https://github.com/goccy/bigquery-emulator) or
    /// drive a wiremock-backed test fixture. Production code should prefer
    /// [`BigQuerySource::new`], which handles credential loading.
    #[doc(hidden)]
    pub fn from_parts(config: BigQuerySourceConfig, client: Client) -> Self {
        Self { config, client }
    }

    /// Resolve the final SQL statement and ordered bind values for a given
    /// parent-record context.
    fn resolve_query(&self, context: &HashMap<String, Value>) -> (String, Vec<Value>) {
        let mut bindings = self.config.params.clone();
        let (rewritten, context_values) = if context.is_empty() {
            (self.config.query.clone(), Vec::new())
        } else {
            substitute_context_bind_params(&self.config.query, context, bindings.len() + 1, |_| {
                "?".to_string()
            })
        };
        bindings.extend(context_values);
        (rewritten, bindings)
    }

    fn build_query_request(&self, query: String, bindings: &[Value]) -> QueryRequest {
        build_query_request(&self.config, query, bindings)
    }

    /// [`Source::discover`] with explicit caps — split out so tests can
    /// exercise the truncation branches without mocking hundreds of tables.
    /// Production code goes through [`Source::discover`], which applies
    /// `MAX_DISCOVER_TABLES` / `MAX_DISCOVER_SCHEMA_FETCHES`.
    #[doc(hidden)]
    pub async fn discover_with_caps(
        &self,
        max_tables: usize,
        max_schema_fetches: usize,
    ) -> Result<Vec<DatasetDescriptor>, FaucetError> {
        let project = &self.config.project_id;
        let discovery_err = |e: gcp_bigquery_client::error::BQError| -> FaucetError {
            FaucetError::Source(format!("bigquery: catalog discovery failed: {e}"))
        };

        // 1. Enumerate every dataset in the project (paged).
        let mut dataset_ids: Vec<String> = Vec::new();
        let mut page_token: Option<String> = None;
        loop {
            let mut opts = DatasetListOptions::default();
            if let Some(t) = page_token.take() {
                opts = opts.page_token(t);
            }
            let resp = self
                .client
                .dataset()
                .list(project, opts)
                .await
                .map_err(discovery_err)?;
            dataset_ids.extend(
                resp.datasets
                    .iter()
                    .map(|d| d.dataset_reference.dataset_id.clone()),
            );
            page_token = resp.next_page_token;
            if page_token.is_none() {
                break;
            }
        }

        // 2. Enumerate physical tables per dataset (paged), capped in total.
        let mut refs: Vec<(String, String)> = Vec::new();
        let mut truncated = false;
        'datasets: for dataset_id in &dataset_ids {
            let mut page_token: Option<String> = None;
            loop {
                let mut opts = TableListOptions::default();
                if let Some(t) = page_token.take() {
                    opts = opts.page_token(t);
                }
                let resp = self
                    .client
                    .table()
                    .list(project, dataset_id, opts)
                    .await
                    .map_err(discovery_err)?;
                for table in resp.tables.unwrap_or_default() {
                    // Physical tables only — views / materialized views /
                    // external tables are not `SELECT *`-scannable datasets in
                    // the same cheap sense. A missing `type` is treated as a
                    // table (BigQuery always sets it in practice).
                    if let Some(kind) = table.r#type.as_deref()
                        && !kind.eq_ignore_ascii_case("TABLE")
                    {
                        continue;
                    }
                    if refs.len() >= max_tables {
                        truncated = true;
                        break 'datasets;
                    }
                    refs.push((dataset_id.clone(), table.table_reference.table_id));
                }
                page_token = resp.next_page_token;
                if page_token.is_none() {
                    break;
                }
            }
        }
        if truncated {
            tracing::warn!(
                cap = max_tables,
                "BigQuery discovery hit the {max_tables}-table cap; remaining tables were not enumerated",
            );
        }
        if refs.len() > max_schema_fetches {
            tracing::warn!(
                cap = max_schema_fetches,
                total = refs.len(),
                "BigQuery discovery found more than {max_schema_fetches} tables; \
                 only the first {max_schema_fetches} get a schema / row estimate",
            );
        }

        // 3. Fetch schema + row count for the first `max_schema_fetches`
        //    tables; the rest are emitted name-only.
        let mut out = Vec::with_capacity(refs.len());
        for (i, (dataset_id, table_id)) in refs.iter().enumerate() {
            if i < max_schema_fetches {
                let table = self
                    .client
                    .table()
                    .get(project, dataset_id, table_id, None)
                    .await
                    .map_err(discovery_err)?;
                let fields = table.schema.fields.unwrap_or_default();
                out.push(table_descriptor(
                    project,
                    dataset_id,
                    table_id,
                    Some(&fields),
                    table.num_rows.as_deref(),
                ));
            } else {
                out.push(table_descriptor(project, dataset_id, table_id, None, None));
            }
        }
        Ok(out)
    }
}

/// Map one BigQuery table-schema field to a JSON-Schema type fragment
/// matching the shape [`faucet_core::schema::infer_schema`] produces.
///
/// `mode: REPEATED` → `array`; `mode: NULLABLE` (BigQuery's default when the
/// mode is omitted) wraps the base type as `["T", "null"]`; `mode: REQUIRED`
/// keeps the bare base type.
fn bq_field_to_json_schema(field: &TableFieldSchema) -> Value {
    let base = match field.r#type {
        FieldType::Integer | FieldType::Int64 => "integer",
        FieldType::Float | FieldType::Float64 | FieldType::Numeric | FieldType::Bignumeric => {
            "number"
        }
        FieldType::Boolean | FieldType::Bool => "boolean",
        FieldType::Record | FieldType::Struct | FieldType::Json => "object",
        // STRING, BYTES, DATE, DATETIME, TIME, TIMESTAMP, GEOGRAPHY,
        // INTERVAL — all serialized as JSON strings by this source.
        _ => "string",
    };
    match field.mode.as_deref() {
        Some(mode) if mode.eq_ignore_ascii_case("REPEATED") => json!({ "type": "array" }),
        Some(mode) if mode.eq_ignore_ascii_case("REQUIRED") => json!({ "type": base }),
        // NULLABLE, or absent (NULLABLE is the BigQuery default).
        _ => faucet_core::nullable_type(json!({ "type": base })),
    }
}

/// Backtick-quote a fully-qualified `project.dataset.table` path for Standard
/// SQL. Backslashes and backticks inside an identifier are escaped (`\\` /
/// `` \` ``) so a hostile identifier cannot break out of the quoted path.
fn bq_quote_path(project: &str, dataset: &str, table: &str) -> String {
    let esc = |s: &str| s.replace('\\', r"\\").replace('`', r"\`");
    format!("`{}.{}.{}`", esc(project), esc(dataset), esc(table))
}

/// Build one [`DatasetDescriptor`] for a BigQuery table. Pure —
/// unit-testable without a live client. `fields`/`num_rows` are `None` for
/// tables past the schema-fetch cap (emitted name-only).
fn table_descriptor(
    project: &str,
    dataset: &str,
    table: &str,
    fields: Option<&[TableFieldSchema]>,
    num_rows: Option<&str>,
) -> DatasetDescriptor {
    let query = format!("SELECT * FROM {}", bq_quote_path(project, dataset, table));
    let mut descriptor = DatasetDescriptor::new(
        format!("{dataset}.{table}"),
        "table",
        json!({ "query": query }),
    );
    if let Some(fields) = fields {
        descriptor = descriptor.with_schema(faucet_core::columns_to_schema(
            fields
                .iter()
                .map(|f| (f.name.clone(), bq_field_to_json_schema(f))),
        ));
    }
    // `numRows` arrives as a decimal string; a missing/unparseable value
    // simply means no estimate.
    if let Some(n) = num_rows.and_then(|s| s.trim().parse::<u64>().ok()) {
        descriptor = descriptor.with_estimated_rows(n);
    }
    descriptor
}

/// Free-standing version of [`BigQuerySource::build_query_request`] — kept
/// separate so unit tests can exercise it without spinning up a real
/// `gcp_bigquery_client::Client`.
fn build_query_request(
    cfg: &BigQuerySourceConfig,
    query: String,
    bindings: &[Value],
) -> QueryRequest {
    let mut req = QueryRequest::new(query);
    req.use_legacy_sql = cfg.use_legacy_sql;
    req.timeout_ms = Some(clamp_timeout_ms(cfg.statement_timeout));
    req.max_results = Some(cfg.max_results_per_page);
    if let Some(location) = &cfg.location {
        req.location = Some(location.clone());
    }

    if !bindings.is_empty() {
        req.parameter_mode = Some("POSITIONAL".to_string());
        req.query_parameters = Some(
            bindings
                .iter()
                .map(|v| QueryParameter {
                    name: None,
                    parameter_type: Some(QueryParameterType {
                        r#type: bq_param_type(v).to_string(),
                        array_type: None,
                        struct_types: None,
                    }),
                    parameter_value: Some(QueryParameterValue {
                        // BigQuery REST always carries the value as a string;
                        // the parameter_type tells the engine how to parse it.
                        // A JSON null becomes a typed NULL (value omitted).
                        value: match v {
                            Value::Null => None,
                            other => Some(stringify_param(other)),
                        },
                        array_values: None,
                        struct_values: None,
                    }),
                })
                .collect(),
        );
    }

    req
}

/// Infer the BigQuery positional-parameter type from the JSON value, so a
/// numeric or boolean bind compares correctly against a numeric/bool column
/// instead of being forced to STRING (#78/#34). Arrays / objects / null fall
/// back to STRING (stringified JSON).
fn bq_param_type(v: &Value) -> &'static str {
    match v {
        Value::Bool(_) => "BOOL",
        Value::Number(n) => {
            if n.is_i64() || n.is_u64() {
                "INT64"
            } else {
                "FLOAT64"
            }
        }
        _ => "STRING",
    }
}

fn stringify_param(v: &Value) -> String {
    match v {
        Value::String(s) => s.clone(),
        other => other.to_string(),
    }
}

fn clamp_timeout_ms(timeout: Duration) -> i32 {
    let ms = timeout.as_millis();
    if ms > i32::MAX as u128 {
        i32::MAX
    } else {
        ms as i32
    }
}

fn schema_fields(qr: &QueryResponse) -> Vec<TableFieldSchema> {
    qr.schema
        .as_ref()
        .and_then(|s| s.fields.clone())
        .unwrap_or_default()
}

fn job_reference(qr: &QueryResponse) -> Result<(String, Option<String>), FaucetError> {
    let r = qr.job_reference.as_ref().ok_or_else(|| {
        FaucetError::Source("BigQuery query response missing jobReference".into())
    })?;
    let job_id = r
        .job_id
        .clone()
        .ok_or_else(|| FaucetError::Source("BigQuery jobReference missing jobId".into()))?;
    Ok((job_id, r.location.clone()))
}

#[async_trait]
impl faucet_core::Source for BigQuerySource {
    fn connector_name(&self) -> &'static str {
        "bigquery"
    }

    fn config_schema(&self) -> Value {
        serde_json::to_value(faucet_core::schema_for!(BigQuerySourceConfig))
            .expect("schema serialization")
    }

    fn dataset_uri(&self) -> String {
        format!(
            "bigquery://{}?query={}",
            self.config.project_id, self.config.query
        )
    }

    fn supports_discover(&self) -> bool {
        true
    }

    /// Enumerate every physical table in the project, with column types from
    /// `tables.get` schemas and a row estimate from `numRows` (catalog
    /// metadata only — no data scan, no billed query). Datasets are listed
    /// via `datasets.list`, tables per dataset via `tables.list` (both
    /// paged); enumeration is capped at `MAX_DISCOVER_TABLES` tables and
    /// per-table schema fetches at `MAX_DISCOVER_SCHEMA_FETCHES` (tables
    /// past that cap are emitted without a schema / estimate).
    async fn discover(&self) -> Result<Vec<DatasetDescriptor>, FaucetError> {
        self.discover_with_caps(MAX_DISCOVER_TABLES, MAX_DISCOVER_SCHEMA_FETCHES)
            .await
    }

    /// Preflight probe for `faucet doctor`. Overrides the default (which pulls a
    /// page via `stream_pages` and would run the configured query — a **billed**
    /// execution). Instead submits the same query with `dryRun: true`, which
    /// validates auth, SQL syntax, and table/permission access **without
    /// executing or billing** it.
    async fn check(
        &self,
        ctx: &faucet_core::check::CheckContext,
    ) -> Result<faucet_core::check::CheckReport, FaucetError> {
        use faucet_core::check::{CheckReport, Probe};
        let start = std::time::Instant::now();
        let mut req =
            build_query_request(&self.config, self.config.query.clone(), &self.config.params);
        req.dry_run = Some(true);

        let probe = async {
            match self.client.job().query(&self.config.project_id, req).await {
                Ok(_) => Ok::<Probe, Probe>(Probe::pass("query", start.elapsed())),
                Err(e) => Err(Probe::fail_hint(
                    "query",
                    start.elapsed(),
                    format!("BigQuery dry-run failed: {e}"),
                    "verify credentials, project_id, dataset/table access, and the SQL",
                )),
            }
        };
        let probe = match tokio::time::timeout(ctx.timeout, probe).await {
            Ok(Ok(p)) | Ok(Err(p)) => p,
            Err(_elapsed) => Probe::fail_hint(
                "query",
                start.elapsed(),
                "BigQuery dry-run timed out",
                "BigQuery did not respond within the check timeout",
            ),
        };
        Ok(CheckReport::single(probe))
    }

    async fn fetch_with_context(
        &self,
        context: &HashMap<String, Value>,
    ) -> Result<Vec<Value>, FaucetError> {
        let (query, bindings) = self.resolve_query(context);
        let req = self.build_query_request(query, &bindings);

        let initial = self
            .client
            .job()
            .query(&self.config.project_id, req)
            .await
            .map_err(|e| FaucetError::Source(format!("BigQuery jobs.query failed: {e}")))?;

        let fields = schema_fields(&initial);
        let mut all_rows: Vec<Value> = rows_from_response(&initial, &fields);
        let mut page_token = initial.page_token.clone();
        let mut job_complete = initial.job_complete.unwrap_or(false);
        let (job_id, job_location) = job_reference(&initial)?;
        let mut fields = fields;
        let poll_timeout = self.config.poll_timeout;
        let poll_started = std::time::Instant::now();

        // Either keep polling until jobComplete, or keep paging until
        // pageToken vanishes. The two reasons we'd loop again share one
        // condition: we are not done.
        while !job_complete || page_token.is_some() {
            let params = GetQueryResultsParameters {
                page_token: page_token.clone(),
                max_results: Some(self.config.max_results_per_page),
                location: job_location.clone(),
                ..Default::default()
            };

            let resp = self
                .client
                .job()
                .get_query_results(&self.config.project_id, &job_id, params)
                .await
                .map_err(|e| {
                    FaucetError::Source(format!("BigQuery jobs.getQueryResults failed: {e}"))
                })?;

            job_complete = resp.job_complete.unwrap_or(false);
            if !job_complete {
                // `poll_timeout == 0` disables the cap (poll forever).
                if !poll_timeout.is_zero() && poll_started.elapsed() >= poll_timeout {
                    return Err(FaucetError::Source(format!(
                        "BigQuery job '{job_id}' did not complete within poll_timeout ({}s)",
                        poll_timeout.as_secs()
                    )));
                }
                tokio::time::sleep(Duration::from_millis(200)).await;
                continue;
            }

            // Fill in the schema from the first complete page if `jobs.query`
            // returned 200 without one (happens when the statement timeout
            // fires before completion).
            if fields.is_empty()
                && let Some(s) = resp.schema.as_ref()
                && let Some(f) = s.fields.as_ref()
            {
                fields = f.clone();
            }

            for row in resp.rows.unwrap_or_default() {
                all_rows.push(row_to_json(&row, &fields));
            }
            page_token = resp.page_token;
            if page_token.is_none() {
                break;
            }
        }

        tracing::info!(
            rows = all_rows.len(),
            query = %self.config.query,
            "BigQuery source fetch complete",
        );
        Ok(all_rows)
    }

    /// Stream rows page-by-page via `jobs.getQueryResults` without
    /// buffering the full result set.
    ///
    /// The trait-level `batch_size` argument is ignored in favour of the
    /// config field — the config is the user-facing knob the README
    /// documents, and routing the pipeline-supplied hint through it would
    /// silently override an explicit config value.
    ///
    /// `batch_size = 0` is the "no batching" sentinel: all rows from all
    /// pages are concatenated and emitted as a single page. The source has
    /// no incremental-replication mode today, so every emitted page carries
    /// `bookmark: None`.
    fn stream_pages<'a>(
        &'a self,
        context: &'a HashMap<String, Value>,
        _batch_size: usize,
    ) -> Pin<Box<dyn Stream<Item = Result<StreamPage, FaucetError>> + Send + 'a>> {
        let batch_size = self.config.batch_size;

        Box::pin(async_stream::try_stream! {
            let (query, bindings) = self.resolve_query(context);
            let req = self.build_query_request(query, &bindings);

            let initial = self
                .client
                .job()
                .query(&self.config.project_id, req)
                .await
                .map_err(|e| FaucetError::Source(format!("BigQuery jobs.query failed: {e}")))?;

            let mut fields = schema_fields(&initial);
            let mut buffer: Vec<Value> = if batch_size == 0 {
                Vec::with_capacity(1024)
            } else {
                Vec::with_capacity(batch_size)
            };
            let chunk = if batch_size == 0 { usize::MAX } else { batch_size };

            for row in rows_from_response_owned(&initial, &fields) {
                buffer.push(row);
                if buffer.len() >= chunk {
                    let page = std::mem::replace(&mut buffer, Vec::with_capacity(chunk));
                    yield StreamPage { records: page, bookmark: None };
                }
            }

            let mut job_complete = initial.job_complete.unwrap_or(false);
            let mut page_token = initial.page_token.clone();

            // If the first response was incomplete, we have to know the job id
            // to keep polling. If it was complete but had no further token,
            // we're done after emitting the first batch.
            let (job_id, job_location) = job_reference(&initial)?;
            let poll_timeout = self.config.poll_timeout;
            let poll_started = std::time::Instant::now();

            while !job_complete || page_token.is_some() {
                let params = GetQueryResultsParameters {
                    page_token: page_token.clone(),
                    max_results: Some(self.config.max_results_per_page),
                    location: job_location.clone(),
                    ..Default::default()
                };

                let resp = self
                    .client
                    .job()
                    .get_query_results(&self.config.project_id, &job_id, params)
                    .await
                    .map_err(|e| {
                        FaucetError::Source(format!("BigQuery jobs.getQueryResults failed: {e}"))
                    })?;

                job_complete = resp.job_complete.unwrap_or(false);
                if !job_complete {
                    // `poll_timeout == 0` disables the cap (poll forever).
                    if !poll_timeout.is_zero() && poll_started.elapsed() >= poll_timeout {
                        Err(FaucetError::Source(format!(
                            "BigQuery job '{job_id}' did not complete within poll_timeout ({}s)",
                            poll_timeout.as_secs()
                        )))?;
                    }
                    tokio::time::sleep(Duration::from_millis(200)).await;
                    continue;
                }

                if fields.is_empty()
                    && let Some(s) = resp.schema.as_ref()
                    && let Some(f) = s.fields.as_ref()
                {
                    fields = f.clone();
                }

                for row in resp.rows.unwrap_or_default() {
                    buffer.push(row_to_json(&row, &fields));
                    if buffer.len() >= chunk {
                        let page = std::mem::replace(&mut buffer, Vec::with_capacity(chunk));
                        yield StreamPage { records: page, bookmark: None };
                    }
                }
                page_token = resp.page_token;
                if page_token.is_none() {
                    break;
                }
            }

            if !buffer.is_empty() {
                yield StreamPage { records: buffer, bookmark: None };
            }

            tracing::info!(
                batch_size,
                query = %self.config.query,
                "BigQuery source stream complete",
            );
        })
    }
}

/// Borrow-based row extraction (used by `fetch_with_context`, which collects
/// into a `Vec` anyway).
fn rows_from_response(resp: &QueryResponse, fields: &[TableFieldSchema]) -> Vec<Value> {
    resp.rows
        .as_ref()
        .map(|rows| rows.iter().map(|r| row_to_json(r, fields)).collect())
        .unwrap_or_default()
}

/// Owned-iteration variant — clones each row out of the response so the
/// streaming loop above doesn't have to keep a borrow open across yields.
fn rows_from_response_owned(resp: &QueryResponse, fields: &[TableFieldSchema]) -> Vec<Value> {
    let rows: &Vec<TableRow> = match resp.rows.as_ref() {
        Some(r) => r,
        None => return Vec::new(),
    };
    rows.iter().map(|r| row_to_json(r, fields)).collect()
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::config::BigQueryCredentials;
    use serde_json::json;

    fn cfg() -> BigQuerySourceConfig {
        BigQuerySourceConfig::new(
            "my-project",
            BigQueryCredentials::ApplicationDefault,
            "SELECT id FROM events",
        )
    }

    #[test]
    fn dataset_uri_returns_project_and_query() {
        // Inline logic test — BigQuerySource::new requires a live client, so
        // we replicate the dataset_uri() computation directly from config fields.
        let c = cfg();
        let uri = format!("bigquery://{}?query={}", c.project_id, c.query);
        assert_eq!(uri, "bigquery://my-project?query=SELECT id FROM events");
    }

    #[test]
    fn stringify_param_passes_strings_unquoted() {
        assert_eq!(stringify_param(&json!("us-east")), "us-east");
        assert_eq!(stringify_param(&json!(42)), "42");
        assert_eq!(stringify_param(&json!(true)), "true");
    }

    #[test]
    fn clamp_timeout_ms_handles_overflow() {
        assert_eq!(clamp_timeout_ms(Duration::from_secs(1)), 1000);
        assert_eq!(clamp_timeout_ms(Duration::from_secs(u64::MAX)), i32::MAX);
    }

    #[test]
    fn build_request_no_params_omits_query_parameters() {
        let c = cfg();
        let req = build_query_request(&c, "SELECT id".to_string(), &[]);
        assert_eq!(req.query, "SELECT id");
        assert!(req.query_parameters.is_none());
        assert!(req.parameter_mode.is_none());
        assert!(!req.use_legacy_sql);
        assert_eq!(req.max_results, Some(1000));
    }

    #[test]
    fn doctor_probe_request_is_dry_run() {
        // The `faucet doctor` `check()` probe submits the configured query with
        // dryRun=true so it validates auth/SQL/permissions without a billed
        // execution — mirror that construction here to guard the field name.
        let c = cfg();
        let mut req = build_query_request(&c, "SELECT 1".to_string(), &[]);
        req.dry_run = Some(true);
        assert_eq!(
            req.dry_run,
            Some(true),
            "doctor probe must dry-run (no billing)"
        );
    }

    #[test]
    fn build_request_with_params_uses_positional_string_binds() {
        let c = cfg().with_params(vec![json!("us-east"), json!(42)]);
        let req = build_query_request(&c, "SELECT * WHERE r = ? AND n > ?".to_string(), &c.params);
        assert_eq!(req.parameter_mode.as_deref(), Some("POSITIONAL"));
        let params = req.query_parameters.as_ref().unwrap();
        assert_eq!(params.len(), 2);
        assert_eq!(params[0].parameter_type.as_ref().unwrap().r#type, "STRING");
        assert_eq!(
            params[0].parameter_value.as_ref().unwrap().value.as_deref(),
            Some("us-east")
        );
        assert_eq!(
            params[1].parameter_value.as_ref().unwrap().value.as_deref(),
            Some("42")
        );
    }

    #[test]
    fn build_request_propagates_location_and_legacy_flag() {
        let c = cfg()
            .with_location("EU")
            .with_use_legacy_sql(true)
            .with_max_results_per_page(250);
        let req = build_query_request(&c, "SELECT 1".to_string(), &[]);
        assert!(req.use_legacy_sql);
        assert_eq!(req.location.as_deref(), Some("EU"));
        assert_eq!(req.max_results, Some(250));
    }

    #[tokio::test]
    async fn new_rejects_out_of_range_batch_size() {
        let mut config = BigQuerySourceConfig::new(
            "my-project",
            BigQueryCredentials::ApplicationDefault,
            "SELECT id FROM events",
        );
        config.batch_size = faucet_core::MAX_BATCH_SIZE + 1;
        match BigQuerySource::new(config).await {
            Err(faucet_core::FaucetError::Config(m)) => {
                assert!(m.contains("batch_size"), "got: {m}")
            }
            _ => panic!("expected a batch_size Config error"),
        }
    }

    // ── discover: pure descriptor-building helpers ───────────────────────────

    fn field(name: &str, ty: FieldType, mode: Option<&str>) -> TableFieldSchema {
        let mut f = TableFieldSchema::new(name, ty);
        f.mode = mode.map(str::to_owned);
        f
    }

    #[test]
    fn bq_field_types_map_to_json_types() {
        for (ty, want) in [
            (FieldType::Integer, "integer"),
            (FieldType::Int64, "integer"),
            (FieldType::Float, "number"),
            (FieldType::Float64, "number"),
            (FieldType::Numeric, "number"),
            (FieldType::Bignumeric, "number"),
            (FieldType::Boolean, "boolean"),
            (FieldType::Bool, "boolean"),
            (FieldType::Record, "object"),
            (FieldType::Struct, "object"),
            (FieldType::Json, "object"),
            (FieldType::String, "string"),
            (FieldType::Bytes, "string"),
            (FieldType::Date, "string"),
            (FieldType::Datetime, "string"),
            (FieldType::Time, "string"),
            (FieldType::Timestamp, "string"),
            (FieldType::Geography, "string"),
            (FieldType::Interval, "string"),
        ] {
            let f = field("c", ty.clone(), Some("REQUIRED"));
            assert_eq!(
                bq_field_to_json_schema(&f),
                json!({ "type": want }),
                "for BigQuery type {ty:?}"
            );
        }
    }

    #[test]
    fn bq_field_mode_nullable_and_absent_wrap_as_nullable() {
        // Explicit NULLABLE and an absent mode (BigQuery's default) both wrap.
        for mode in [Some("NULLABLE"), None] {
            let f = field("c", FieldType::Integer, mode);
            assert_eq!(
                bq_field_to_json_schema(&f),
                json!({ "type": ["integer", "null"] }),
                "for mode {mode:?}"
            );
        }
    }

    #[test]
    fn bq_field_mode_repeated_maps_to_array() {
        let f = field("tags", FieldType::String, Some("REPEATED"));
        assert_eq!(bq_field_to_json_schema(&f), json!({ "type": "array" }));
    }

    #[test]
    fn bq_quote_path_backtick_quotes_and_escapes() {
        assert_eq!(
            bq_quote_path("proj", "sales", "orders"),
            "`proj.sales.orders`"
        );
        // A hostile identifier cannot break out of the quoted path.
        assert_eq!(bq_quote_path("p", "d", r"we`ird\x"), r"`p.d.we\`ird\\x`");
    }

    #[test]
    fn table_descriptor_carries_schema_estimate_and_patch() {
        let fields = vec![
            field("id", FieldType::Integer, Some("REQUIRED")),
            field("note", FieldType::String, Some("NULLABLE")),
        ];
        let d = table_descriptor("proj", "sales", "orders", Some(&fields), Some("120"));
        assert_eq!(d.name, "sales.orders");
        assert_eq!(d.kind, "table");
        assert_eq!(d.estimated_rows, Some(120));
        assert_eq!(d.config_patch["query"], "SELECT * FROM `proj.sales.orders`");
        let schema = d.schema.as_ref().unwrap();
        assert_eq!(schema["type"], "object");
        assert_eq!(schema["properties"]["id"]["type"], "integer");
        assert_eq!(
            schema["properties"]["note"]["type"],
            json!(["string", "null"])
        );
    }

    #[test]
    fn table_descriptor_without_schema_fetch_is_name_only() {
        // Past the schema-fetch cap: still a full config_patch, no schema/rows.
        let d = table_descriptor("proj", "ops", "events", None, None);
        assert_eq!(d.name, "ops.events");
        assert!(d.schema.is_none());
        assert_eq!(d.estimated_rows, None);
        assert_eq!(d.config_patch["query"], "SELECT * FROM `proj.ops.events`");
    }

    #[test]
    fn table_descriptor_unparseable_num_rows_means_no_estimate() {
        let d = table_descriptor("p", "d", "t", Some(&[]), Some("not-a-number"));
        assert_eq!(d.estimated_rows, None);
        // An empty schema fetch still yields an (empty) object schema.
        assert_eq!(d.schema.as_ref().unwrap()["type"], "object");
    }

    #[test]
    fn resolve_query_substitutes_context_with_positional_markers() {
        // Test resolve_query without needing a Client by mimicking its core.
        let c = cfg();
        let mut bindings = c.params.clone();
        let mut ctx = HashMap::new();
        ctx.insert("parent.id".to_string(), json!(7));
        let (rewritten, extra) = substitute_context_bind_params(
            "SELECT * FROM t WHERE id = {parent.id}",
            &ctx,
            bindings.len() + 1,
            |_| "?".to_string(),
        );
        bindings.extend(extra);
        assert_eq!(rewritten, "SELECT * FROM t WHERE id = ?");
        assert_eq!(bindings, vec![json!(7)]);
    }
}