job-search/internal/findatender/api.go

package findatender

import (
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"log/slog"
	"regexp"
	"strconv"
	"strings"
	"time"

	resty "resty.dev/v3"

	"git.avsdev.uk/Go/job-search/internal/findatender/ocds"
	"git.avsdev.uk/Go/job-search/internal/logging"
)

// Errors returned when fetching from the API.
var (
	// ErrNoNext means a release package has no next page link.
	ErrNoNext = errors.New("no next link")
	// ErrNoReleases means the API returned no releases.
	ErrNoReleases = errors.New("no releases")
	// ErrRateLimitExceeded means the API refused a request with 429 Too Many
	// Requests, and waiting for it was not allowed or did not help.
	ErrRateLimitExceeded = errors.New("rate limit has been exceeded")
	// ErrFetchCancelled means the fetch's context was cancelled while waiting
	// out a rate limit.
	ErrFetchCancelled = errors.New("fetch cancelled")
	// ErrUnexpectedStatus wraps an API response status other than 200 OK.
	ErrUnexpectedStatus = errors.New("unexpected response status")
)

// RequestOptions selects the releases a fetch asks for.
type RequestOptions struct {
	// Since asks for releases updated from this time. Zero means 24 hours ago.
	Since time.Time
	// Limit is the number of releases per page. Zero means 100.
	Limit int
	// AllowRateLimit waits out a 429 response and retries once, instead of
	// failing with ErrRateLimitExceeded.
	AllowRateLimit bool
}

// ---------- Utility methods for handling API irregularities -----------

// safeReleasePackage is ocds.ReleasePackage with the releases left undecoded,
// so safeUnpackReleases can repair or skip each one.
type safeReleasePackage struct {
	URI               string                        `json:"uri,omitempty"`
	Version           string                        `json:"version,omitempty"`
	Extensions        []string                      `json:"extensions,omitempty"`
	PublishedDate     *string                       `json:"publishedDate,omitempty"`
	Releases          []json.RawMessage             `json:"releases,omitempty"`
	Publisher         *ocds.ReleasePackagePublisher `json:"publisher,omitempty"`
	License           string                        `json:"license,omitempty"`
	PublicationPolicy string                        `json:"publicationPolicy,omitempty"`
	Links             *ocds.ReleasePackageLinks     `json:"links,omitempty"`
}

// dateOnly matches a calendar date with no time of day: "2006-10-02".
var dateOnly = regexp.MustCompile(`^\d{4}-\d{2}-\d{2}$`)

// unmarshalLenientDates decodes raw json into v like json.Unmarshal, after
// repairing date fields that time.Time cannot parse.
func unmarshalLenientDates(raw json.RawMessage, v any) error {
	fixed, err := repairDates(raw)
	if err != nil {
		return err
	}
	return json.Unmarshal(fixed, v)
}

// repairDates returns a copy of raw in which every date field holding a
// date-only value ("2006-10-02") is rewritten as midnight UTC
// ("2006-10-02T00:00:00Z"), and every date field holding an empty string is
// set to null. Values that already have a time, values in other fields, and
// dates that appear inside free text are left alone. If nothing needs
// repairing, raw is returned unchanged.
//
// A date field is one whose name is "date" or starts with "date", contains
// "Date", or ends with "Deadline".
func repairDates(raw json.RawMessage) (json.RawMessage, error) {
	dec := json.NewDecoder(bytes.NewReader(raw))
	dec.UseNumber() // keep numbers exactly as written, not as float64

	var doc any
	if err := dec.Decode(&doc); err != nil {
		return nil, fmt.Errorf("repairing dates: %w", err)
	}

	if _, err := dec.Token(); !errors.Is(err, io.EOF) {
		return nil, errors.New(
			"repairing dates: unexpected data after the JSON value",
		)
	}

	if !recursiveRepairDates(doc) {
		return raw, nil
	}

	return json.Marshal(doc)
}

// recursiveRepairDates fixes date fields in v in place and reports whether it
// changed anything.
func recursiveRepairDates(v any) bool {
	changed := false
	switch v := v.(type) {
	case map[string]any:
		for key, val := range v {
			if s, ok := val.(string); ok && isDateField(key) {
				switch {
				case dateOnly.MatchString(s):
					v[key] = s + "T00:00:00Z"
					changed = true
				case s == "":
					v[key] = nil
					changed = true
				}
				continue
			}
			if recursiveRepairDates(val) {
				changed = true
			}
		}
	case []any:
		for _, el := range v {
			if recursiveRepairDates(el) {
				changed = true
			}
		}
	}
	return changed
}

// isDateField reports whether a field holds a date, by its name (see
// repairDates).
func isDateField(name string) bool {
	return strings.HasPrefix(name, "date") ||
		strings.Contains(name, "Date") ||
		strings.HasSuffix(name, "Deadline")
}

// releaseIDs reads just the ocid and the release id from a release, without
// decoding the rest of it. It reads the top-level fields in order and stops
// as soon as it has both, so an error later in the release (invalid JSON or a
// value that wouldn't fit the release struct) doesn't matter. Fields before
// them only need to be valid JSON.
func releaseIDs(raw json.RawMessage) (ocid, id string, err error) {
	dec := json.NewDecoder(bytes.NewReader(raw))
	dec.UseNumber() // a numeric id keeps its exact digits

	if tok, err := dec.Token(); err != nil {
		return "", "", fmt.Errorf("reading release: %w", err)
	} else if tok != json.Delim('{') {
		return "", "", errors.New("reading release: not a JSON object")
	}

	var haveOCID, haveID bool
	for dec.More() {
		tok, err := dec.Token()
		if err != nil {
			return ocid, id, fmt.Errorf("reading release: %w", err)
		}
		switch key, _ := tok.(string); key {
		case "ocid":
			if err := dec.Decode(&ocid); err != nil {
				return ocid, id, fmt.Errorf("reading release ocid: %w", err)
			}
			haveOCID = true
		case "id":
			if id, err = decodeReleaseID(dec); err != nil {
				return ocid, id, err
			}
			haveID = true
		default:
			var skip json.RawMessage // must be valid JSON, but is otherwise ignored
			if err := dec.Decode(&skip); err != nil {
				return ocid, id, fmt.Errorf("reading release field %q: %w", key, err)
			}
		}
		if haveOCID && haveID {
			return ocid, id, nil
		}
	}

	return ocid, id, missingIDsError(haveOCID, haveID)
}

// decodeReleaseID decodes a release id. OCDS ids are strings, but a number is
// tolerated and kept as written.
func decodeReleaseID(dec *json.Decoder) (string, error) {
	var v any
	if err := dec.Decode(&v); err != nil {
		return "", fmt.Errorf("reading release id: %w", err)
	}
	switch v := v.(type) {
	case string:
		return v, nil
	case json.Number:
		return v.String(), nil
	default:
		return "", fmt.Errorf("reading release id: unexpected %T", v)
	}
}

// missingIDsError reports which of the ocid and release id were not found.
func missingIDsError(haveOCID, haveID bool) error {
	switch {
	case !haveOCID && !haveID:
		return errors.New("reading release: no ocid or id")
	case !haveOCID:
		return errors.New("reading release: no ocid")
	default:
		return errors.New("reading release: no id")
	}
}

// safeUnpackReleases unpacks an API response into an OCDS ReleasePackage,
// handling API irregularities (malformed dates, missing fields, invalid
// releases, etc).
//
// This method takes ownership of the response body and closes it when done.
func safeUnpackReleases(response *resty.Response) (*ocds.ReleasePackage, error) {
	var pkg safeReleasePackage

	// Take ownership of the response body.
	defer response.Body.Close() //nolint:errcheck // read only

	decoder := json.NewDecoder(response.Body)
	if err := decoder.Decode(&pkg); err != nil {
		return nil, err
	}

	out := &ocds.ReleasePackage{
		URI:               pkg.URI,
		Version:           pkg.Version,
		Extensions:        pkg.Extensions,
		Publisher:         pkg.Publisher,
		License:           pkg.License,
		PublicationPolicy: pkg.PublicationPolicy,
		Links:             pkg.Links,
	}

	// Correct PublishedDate of "" and "2006-01-02"
	if pkg.PublishedDate != nil && len(*pkg.PublishedDate) > 0 {
		t, err := time.Parse(time.RFC3339, *pkg.PublishedDate)
		if err != nil {
			t, err = time.Parse(time.DateOnly, *pkg.PublishedDate)
		}
		if err == nil {
			out.PublishedDate = &t
		}
	}

	for _, rawRel := range pkg.Releases {
		var release ocds.Release
		err := unmarshalLenientDates(rawRel, &release)
		if err != nil {
			ocid, release_id, _ := releaseIDs(rawRel)
			slog.Warn(fmt.Sprintf(
				"Skipping release (%s/%s) due to error: %s", ocid, release_id, err,
			))
			continue
		}
		out.Releases = append(out.Releases, release)
	}

	return out, nil
}

// ---------- API interaction methods -----------

// rateSleep sleeps the fetch worker until either the context is cancelled
// (shutdown), returning ctx.Err(), or the duration has expired. The duration
// is usually provided by the API; a zero duration returns immediately.
func rateSleep(ctx context.Context, duration time.Duration) error {
	timer := time.NewTimer(duration)
	defer timer.Stop()

	select {
	case <-timer.C:
		return nil
	case <-ctx.Done():
		return ctx.Err()
	}
}

// closeBody closes res's body if it has one, avoiding a double close.
func closeBody(res *resty.Response) {
	if res.Body != nil {
		res.Body.Close() //nolint:errcheck,gosec // read only
	}
}

// fetchURI performs a single GET of uri, using ft's context, and traces it at
// debug level. Unlike the method of the same name it does not check the
// response status.
func fetchURI(ft *FindATender, uri string) (*resty.Response, error) {
	withTrace := ft.config.Logging.LogLevel() <= slog.LevelDebug

	slog.Info("findTender.resty", "request_uri", uri)

	ctx := ft.ctx
	if ctx == nil {
		ctx = context.Background()
	}

	req := ft.httpClient.R()
	req.SetContext(ctx)
	res, err := req.SetTrace(withTrace).Get(uri)

	if withTrace {
		if logging.IsJSONHandler(slog.Default().Handler()) {
			slog.Debug(
				"Resty",
				"trace", res.Request.TraceInfo().JSON(),
				"status", res.Status(),
				"response", res,
			)
		} else {
			slog.Debug("Resty " + res.Request.TraceInfo().String())
			slog.Debug("Resty status: " + res.Status())
			slog.Debug("Resty response: " + res.String())
		}
	}

	return res, err
}

// fetchURI performs a fetch against the API. If the API returns with a rate limit
// reached response and allowRateLimit is true, then wait and retry like a good
// client. If allowRateLimit is false, return to the caller informing them of
// the rate limit exceeded error. If, after waiting, the API still thinks we
// have exceeded the rate limit, bail out anyway.
func (ft *FindATender) fetchURI(
	uri string,
	allowRateLimit bool,
) (*resty.Response, error) {
	res, err := fetchURI(ft, uri)
	if err != nil {
		return nil, err
	}

	ctx := ft.ctx
	if ctx == nil {
		ctx = context.Background()
	}

	if res.StatusCode() == 429 {
		closeBody(res)

		if !allowRateLimit {
			return nil, ErrRateLimitExceeded
		}

		retrySecStr := res.Header().Get("Retry-After")
		retrySec, err := strconv.ParseInt(retrySecStr, 10, 32)
		if err != nil {
			return nil, fmt.Errorf("unknown retry-after value: %w", err)
		}

		slog.Warn(fmt.Sprintf(
			"API rate limit hit, waiting as requested for %d seconds", retrySec,
		))
		if rateSleep(ctx, time.Duration(retrySec)*time.Second) != nil {
			return nil, ErrFetchCancelled
		}

		res, err = fetchURI(ft, uri)
		if err != nil {
			return nil, err
		}
		if res.StatusCode() == 429 {
			closeBody(res)
			return nil, ErrRateLimitExceeded
		}
	}

	if res.StatusCode() != 200 {
		closeBody(res)
		return nil, fmt.Errorf("%w: %s", ErrUnexpectedStatus, res.Status())
	}

	return res, nil
}

// fetchWithOpts performs a fetch against the API with some sanitised default
// options (see RequestOptions).
func (ft *FindATender) fetchWithOpts(
	opts RequestOptions,
) (*resty.Response, error) {
	if opts.Since.IsZero() {
		opts.Since = time.Now().AddDate(0, 0, -1)
	}
	if opts.Limit == 0 {
		opts.Limit = 100
	}

	uri := fmt.Sprintf(
		"%s/ocdsReleasePackages?updatedFrom=%s&limit=%d",
		ft.apiRoot,
		opts.Since.Format("2006-01-02T15:04:05"),
		opts.Limit,
	)

	return ft.fetchURI(uri, opts.AllowRateLimit)
}

// fetchNext fetches the page pkg's "Next" link points to. It returns
// ErrNoNext if there is no "Next" link; the caller can decide whether asking
// for it was an error.
func (ft *FindATender) fetchNext(
	pkg *ocds.ReleasePackage,
	allowRateLimit bool,
) (*resty.Response, error) {
	if pkg.Links == nil || len(pkg.Links.Next) == 0 {
		return nil, ErrNoNext
	}
	return ft.fetchURI(pkg.Links.Next, allowRateLimit)
}

// fetchOnce makes one call on the API. It does not follow any "Next" links.
func (ft *FindATender) fetchOnce(opts RequestOptions) ([]ocds.Release, error) {
	res, err := ft.fetchWithOpts(opts)
	if err != nil {
		return nil, err
	}

	pkg, err := safeUnpackReleases(res)
	if err != nil {
		return nil, err
	}

	if len(pkg.Releases) == 0 {
		return nil, ErrNoReleases
	}

	return pkg.Releases, nil
}

// fetchAll makes a call on the API. If there are any "Next" links, it calls
// them in succession until the end has been reached.
func (ft *FindATender) fetchAll(opts RequestOptions) ([]ocds.Release, error) {
	var releases []ocds.Release

	res, err := ft.fetchWithOpts(opts)
	if err != nil {
		return nil, err
	}

	pkg, err := safeUnpackReleases(res)
	if err != nil {
		return nil, err
	}
	releases = append(releases, pkg.Releases...)

	// Next links already followed, so a page that links back to an earlier
	// one cannot loop forever.
	seen := make(map[string]bool)
	for {
		if pkg.Links != nil && len(pkg.Links.Next) > 0 {
			if seen[pkg.Links.Next] {
				slog.Warn(fmt.Sprintf(
					"Stopping at repeated next link: %s", pkg.Links.Next,
				))
				break
			}
			seen[pkg.Links.Next] = true
		}

		res, err := ft.fetchNext(pkg, opts.AllowRateLimit)
		if err == ErrNoNext {
			break
		}
		if err != nil {
			return nil, err
		}

		pkg, err = safeUnpackReleases(res)
		if err != nil {
			return nil, err
		}
		releases = append(releases, pkg.Releases...)
	}

	if len(releases) == 0 {
		return nil, ErrNoReleases
	}

	return releases, nil
}