sqlload 0.46.0

SQL data loading tooling.
//! Paging through one endpoint with its named paginator.

use std::collections::HashSet;
use std::collections::hash_map::DefaultHasher;
use std::hash::{Hash, Hasher};
use std::sync::atomic::Ordering;

use reqwest::Url;
use serde_json::{Map, Value};
use tokio::sync::mpsc::Sender;

use crate::http_source::_helpers::cells::pivoted_rows;
use crate::http_source::_helpers::fetch::fetch;
use crate::http_source::_helpers::job::request_projection;
use crate::http_source::_helpers::selector::{
    missing_path, pointer_text, pointer_value, selected_node, selector,
};
use crate::http_source::_helpers::urls::{redacted, request_url};
use crate::http_source::models::{
    Calls, CellNames, Fetched, HttpPage, Location, Method, PageCall, Paginator, Request,
    SelectMissing, Selector,
};
use crate::python_sources::errors::ReadError;

/// Send every page of `request` at `path` (its placeholders filled) to `pages`, with where its rows are; a page whose selected node is null has no rows and is not sent, and one missing it fails, naming what the response holds, unless the request reads a missing node as no rows (`select_missing`). An offset run that would end after a first page shorter than its limit asks for the next offset once, and fails when rows come back, as the API then caps its pages lower than the limit. Stops quietly once `pages` closes; a page read twice, or a response repeating the previous one, is an error rather than a loop.
pub(crate) async fn paginate(
    calls: &Calls,
    request: &Request,
    path: &str,
    pages: &Sender<HttpPage>,
) -> Result<(), ReadError> {
    let rows = selector(request.select.as_deref())?;
    let projection = request_projection(request)?;
    let rows_path = Some(rows.rows.join(".")).filter(|path| !path.is_empty());
    let first = request_url(&calls.endpoint.base_url, path)?;
    let mut call = page_call(request, &first, start_values(&request.paginator))?;
    let mut seen: HashSet<(Url, Option<String>)> = HashSet::new();
    let mut previous: Option<u64> = None;
    let mut names: Option<CellNames> = None;
    for step in 1_u64.. {
        let shown = redacted(
            &format!("{} {}", call.method.as_str(), call.url),
            &calls.endpoint.current_secrets(),
        );
        let body_text = call.body.as_ref().map(Value::to_string);
        if !seen.insert((call.url.clone(), body_text)) {
            return Err(format!(
                "{shown}: the paginator came back to a page it already read; check its pointer or parameter"
            )
            .into());
        }
        let Some(fetched) = fetch(calls, &call).await? else {
            return Ok(());
        };
        let digest = counted(&request.paginator).then(|| digest(&fetched.bytes));
        if digest.is_some() && previous == digest {
            return Err(format!(
                "{shown}: the response repeats the previous page; the API may ignore the paginator's parameter"
            )
            .into());
        }
        previous = digest;
        let document = parsed(request, &rows, &fetched.bytes)
            .map_err(|error| format!("{shown}: the response is not JSON: {error}"))?;
        if let Some(document) = &document {
            selected_or_allowed(request, (&rows, document), &shown)?;
        }
        let next = next_call(
            request,
            (&first, &call),
            (&fetched, document.as_ref(), step),
        )
        .map_err(|error| format!("{shown}: {error}"))?;
        match (step, &next, &document) {
            (1, None, Some(document)) => {
                capped_pages(calls, request, (&first, document), &shown).await?
            }
            (_, Some(_), _) if matches!(request.paginator, Paginator::Offset { .. }) => {
                calls.full_page_seen.store(true, Ordering::Relaxed);
            }
            _ => {}
        }
        let has_rows = document.as_ref().is_none_or(|document| {
            selected_node(document, &rows).is_some_and(|node| !node.is_null())
        });
        let (bytes, path) = match (&request.cells, document.as_ref()) {
            (Some(cells), Some(document)) => {
                let (text, used) = pivoted_rows(document, (&rows, cells), names.take())
                    .map_err(|error| format!("{shown}: cells: {error}"))?;
                names = used;
                (text.into_bytes(), None)
            }
            (Some(_), None) | (None, _) => (fetched.bytes, rows_path.clone()),
        };
        let page = HttpPage {
            bytes,
            path,
            source: shown,
            projection: projection.clone(),
        };
        if has_rows && pages.send(page).await.is_err() {
            return Ok(());
        }
        match next {
            Some(next) => call = next,
            None => return Ok(()),
        }
    }
    Ok(())
}

/// For offset paging whose first page (`document`, from the request `shown`) held some rows but fewer than its limit: the page at the next offset, which an API capping its pages below the limit fills, so rows there fail with the page size to set. An empty or missing page is the real end, and so are the first page's rows again, from an API that rounds the offset down to a page boundary. Skipped once a page of the request held its whole limit in this job, which shows the API does not cap it (see `Calls::full_page_seen`).
async fn capped_pages(
    calls: &Calls,
    request: &Request,
    (first, document): (&Url, &Value),
    shown: &str,
) -> Result<(), ReadError> {
    let Paginator::Offset {
        param,
        limit_param,
        limit,
        ..
    } = &request.paginator
    else {
        return Ok(());
    };
    let read = page_rows(document, request).map_err(|error| format!("{shown}: {error}"))?;
    if read == 0 || read >= *limit || calls.full_page_seen.load(Ordering::Relaxed) {
        return Ok(());
    }
    let probe = page_call(
        request,
        first,
        vec![
            (param.clone(), Value::from(read)),
            (limit_param.clone(), Value::from(*limit)),
        ],
    )?;
    let Some(fetched) = fetch(calls, &probe).await? else {
        return Ok(());
    };
    let probed = redacted(
        &format!("{} {}", probe.method.as_str(), probe.url),
        &calls.endpoint.current_secrets(),
    );
    let rows = selector(request.select.as_deref())?;
    let more = match parsed(request, &rows, &fetched.bytes)
        .map_err(|error| format!("{probed}: the response is not JSON: {error}"))?
    {
        Some(next) if selected_node(&next, &rows) == selected_node(document, &rows) => 0,
        Some(next) => {
            selected_or_allowed(request, (&rows, &next), &probed)?;
            page_rows(&next, request).map_err(|error| format!("{probed}: {error}"))?
        }
        None => 0,
    };
    match more {
        0 => Ok(()),
        _ => Err(format!(
            "{shown}: the first page held {read} rows, fewer than limit {limit}, but the API returned {more} more at offset {read}, so it caps its pages at {read} rows; set limit={read}"
        )
        .into()),
    }
}

/// Fails when the response `document` to `shown` lacks the `rows` path, naming what it holds, unless the request reads a missing path as no rows (`select_missing`).
fn selected_or_allowed(
    request: &Request,
    (rows, document): (&Selector, &Value),
    shown: &str,
) -> Result<(), ReadError> {
    match request.select_missing == SelectMissing::Fail && missing_path(document, rows) {
        true => Err(format!(
            "{shown}: select {:?} is not in the response, which holds {}; an API can answer an error with a 200 status, so check the response. For an API that leaves an empty list out, set select_missing=\"empty\" to read such a page as no rows",
            request.select.clone().unwrap_or_default(),
            top_level(document)
        )
        .into()),
        false => Ok(()),
    }
}

/// What a response holds at its top level, for an error: its keys, or the kind of value it is.
fn top_level(document: &Value) -> String {
    match document {
        Value::Object(fields) if fields.is_empty() => "an empty object".to_owned(),
        Value::Object(fields) => {
            let keys: Vec<&str> = fields.keys().map(String::as_str).collect();
            format!("the keys {}", keys.join(", "))
        }
        Value::Array(_) => "a list".to_owned(),
        Value::String(_) => "text".to_owned(),
        Value::Number(_) => "a number".to_owned(),
        Value::Bool(_) => "true or false".to_owned(),
        Value::Null => "null".to_owned(),
    }
}

/// The response as JSON when the selector or the paginator reads it, else `None`.
fn parsed(request: &Request, rows: &Selector, bytes: &[u8]) -> Result<Option<Value>, String> {
    let reads_body = match &request.paginator {
        Paginator::Single | Paginator::LinkHeader => !rows.rows.is_empty(),
        Paginator::NextLink { .. }
        | Paginator::Cursor { .. }
        | Paginator::Offset { .. }
        | Paginator::PageNumber { .. } => true,
    } || request.cells.is_some();
    match reads_body {
        true => serde_json::from_slice(bytes)
            .map(Some)
            .map_err(|error| error.to_string()),
        false => Ok(None),
    }
}

/// Whether the paginator moves by a counter whatever the response says, so only a repeated body reveals an API ignoring it; the others follow the response, and the seen URLs catch their loops.
fn counted(paginator: &Paginator) -> bool {
    match paginator {
        Paginator::Offset { .. } | Paginator::PageNumber { .. } => true,
        Paginator::Single
        | Paginator::NextLink { .. }
        | Paginator::LinkHeader
        | Paginator::Cursor { .. } => false,
    }
}

fn digest(bytes: &[u8]) -> u64 {
    let mut hasher = DefaultHasher::new();
    bytes.hash(&mut hasher);
    hasher.finish()
}

/// The paginator's values on the first request.
fn start_values(paginator: &Paginator) -> Vec<(String, Value)> {
    match paginator {
        Paginator::Offset {
            param,
            limit_param,
            limit,
            ..
        } => vec![
            (param.clone(), Value::from(0_u64)),
            (limit_param.clone(), Value::from(*limit)),
        ],
        Paginator::PageNumber { param, start, .. } => vec![(param.clone(), Value::from(*start))],
        Paginator::Single
        | Paginator::NextLink { .. }
        | Paginator::LinkHeader
        | Paginator::Cursor { .. } => Vec::new(),
    }
}

/// Where the paginator's values go.
fn location(paginator: &Paginator) -> Location {
    match paginator {
        Paginator::Offset { location, .. }
        | Paginator::PageNumber { location, .. }
        | Paginator::Cursor { location, .. } => *location,
        Paginator::Single | Paginator::NextLink { .. } | Paginator::LinkHeader => Location::Query,
    }
}

/// The request at `url` with the paginator's `values`: in the query after `params`, or at their paths in a fresh copy of `json_body`, which is never changed itself. A `POST` without `json_body` starts from `{}`.
fn page_call(
    request: &Request,
    url: &Url,
    values: Vec<(String, Value)>,
) -> Result<PageCall, String> {
    let (query, in_body) = match location(&request.paginator) {
        Location::Query => (values, Vec::new()),
        Location::Body => (Vec::new(), values),
    };
    let query: Vec<(String, String)> = query
        .into_iter()
        .map(|(name, value)| (name, param_text(&value)))
        .collect();
    let body = match request.method {
        Method::Get => None,
        Method::Post => {
            let mut body = request
                .json_body
                .clone()
                .unwrap_or_else(|| Value::Object(Map::new()));
            for (path, value) in in_body {
                set_at(&mut body, &path, value)?;
            }
            Some(body)
        }
    };
    Ok(PageCall {
        method: request.method,
        url: with_params(url, &request.params, &query),
        body,
    })
}

/// `body` with `value` at `path` (such as `$.page.cursor`), adding objects on the way.
pub(crate) fn set_at(body: &mut Value, path: &str, value: Value) -> Result<(), String> {
    let names = selector(Some(path))?.rows;
    let Some((last, parents)) = names.split_last() else {
        return Err(format!("body path {path:?} names no field"));
    };
    let mut node = body;
    for name in parents {
        node = node
            .as_object_mut()
            .ok_or_else(|| format!("json_body has no object to set {path} in"))?
            .entry(name.clone())
            .or_insert_with(|| Value::Object(Map::new()));
    }
    node.as_object_mut()
        .ok_or_else(|| format!("json_body has no object to set {path} in"))?
        .insert(last.clone(), value);
    Ok(())
}

/// A value as a query parameter: text as itself, anything else as JSON.
fn param_text(value: &Value) -> String {
    match value {
        Value::String(text) => text.clone(),
        other => other.to_string(),
    }
}

/// The request after `fetched` (the `step`th page, sent as `current`), or `None` at the end; links resolve against the URL the response came from, after any redirect, and keep the current body; `document` is the parsed response whenever the paginator reads the body.
fn next_call(
    request: &Request,
    (first, current): (&Url, &PageCall),
    (fetched, document, step): (&Fetched, Option<&Value>, u64),
) -> Result<Option<PageCall>, String> {
    let body = || document.ok_or_else(|| "the paginator needs a JSON response".to_owned());
    let joined = |link: String| {
        fetched
            .url
            .join(&link)
            .map(|url| {
                Some(PageCall {
                    url,
                    ..current.clone()
                })
            })
            .map_err(|error| format!("next page {link} is not a URL: {error}"))
    };
    let next = |values: Vec<(String, Value)>| page_call(request, first, values).map(Some);
    match &request.paginator {
        Paginator::Single => Ok(None),
        Paginator::LinkHeader => fetched.next_link.clone().map_or(Ok(None), joined),
        Paginator::NextLink { pointer } => {
            pointer_text(body()?, &selector(Some(pointer))?).map_or(Ok(None), joined)
        }
        Paginator::Cursor { pointer, param, .. } => {
            match pointer_value(body()?, &selector(Some(pointer))?) {
                Some(token) => next(vec![(param.clone(), token)]),
                None => Ok(None),
            }
        }
        Paginator::Offset {
            param,
            limit_param,
            limit,
            ..
        } => match page_rows(body()?, request)? >= *limit {
            true => next(vec![
                (param.clone(), Value::from(step * limit)),
                (limit_param.clone(), Value::from(*limit)),
            ]),
            false => Ok(None),
        },
        Paginator::PageNumber { param, start, .. } => match page_rows(body()?, request)? > 0 {
            true => next(vec![(param.clone(), Value::from(start + step))]),
            false => Ok(None),
        },
    }
}

/// How many rows a page holds for offset and page-number paging: its selected list's length, 0 when missing or null; anything else is an error, since a counted object would never end the paging.
fn page_rows(document: &Value, request: &Request) -> Result<u64, String> {
    match selected_node(document, &selector(request.select.as_deref())?) {
        Some(Value::Array(items)) => Ok(items.len() as u64),
        Some(Value::Null) | None => Ok(0),
        Some(_) => Err(format!(
            "offset and page-number paging count the rows in a list, but select {:?} is not one; point select at the page's list of rows",
            request.select.clone().unwrap_or_else(|| "$".to_owned())
        )),
    }
}

/// `url` with `params` (nulls left out) and then `extra` appended to its query.
fn with_params(url: &Url, params: &Map<String, Value>, extra: &[(String, String)]) -> Url {
    let mut url = url.clone();
    let pairs = params.iter().filter_map(|(name, value)| match value {
        Value::Null => None,
        Value::String(text) => Some((name.clone(), text.clone())),
        other => Some((name.clone(), other.to_string())),
    });
    let pairs: Vec<(String, String)> = pairs.chain(extra.iter().cloned()).collect();
    if !pairs.is_empty() {
        url.query_pairs_mut().extend_pairs(pairs);
    }
    url
}