Ingest checksum issue (#378)

- Refactor `Downloader.Get`.

CHANGELOG.md

@@ -7,7 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 Breaking changes are annotated with ☢️, and alpha/beta features with 🐥.
 
-## [v0.47.0] - 2024-01-28
+## [v0.47.0] - UPCOMING
 
 This is a significant release, focused on improving i/o, responsiveness,
 and performance. The headline features are [caching](https://sq.io/docs/source#cache)

libsq/core/ioz/ioz.go

@@ -5,6 +5,7 @@ import (
 	"bytes"
 	"context"
 	crand "crypto/rand"
+	"errors"
 	"fmt"
 	"io"
 	mrand "math/rand"
@@ -382,6 +383,87 @@ func (w *notifyOnceWriter) Write(p []byte) (n int, err error) {
 	return w.w.Write(p)
 }
 
+// NotifyOnEOFReader returns an [io.Reader] that invokes fn
+// when r.Read returns [io.EOF]. The error that fn returns is
+// what's returned to the r caller: fn can transform the error
+// or return it unchanged. If r or fn is nil, r is returned.
+//
+// If r is an [io.ReadCloser], the returned reader will also
+// implement [io.ReadCloser].
+//
+// See also: [NotifyOnErrorReader], which is a generalization of
+// [NotifyOnEOFReader].
+func NotifyOnEOFReader(r io.Reader, fn func(error) error) io.Reader {
+	if r == nil || fn == nil {
+		return r
+	}
+
+	if rc, ok := r.(io.ReadCloser); ok {
+		return &notifyOnEOFReadCloser{notifyOnEOFReader{r: rc, fn: fn}}
+	}
+
+	return &notifyOnEOFReader{r: r}
+}
+
+type notifyOnEOFReader struct {
+	r  io.Reader
+	fn func(error) error
+}
+
+// Read implements io.Reader.
+func (r *notifyOnEOFReader) Read(p []byte) (n int, err error) {
+	n, err = r.r.Read(p)
+	if err != nil && errors.Is(err, io.EOF) {
+		err = r.fn(err)
+	}
+
+	return n, err
+}
+
+var _ io.ReadCloser = (*notifyOnEOFReadCloser)(nil)
+
+type notifyOnEOFReadCloser struct {
+	notifyOnEOFReader
+}
+
+// Close implements io.Closer.
+func (r *notifyOnEOFReadCloser) Close() error {
+	if c, ok := r.r.(io.Closer); ok {
+		return c.Close()
+	}
+	return nil
+}
+
+// NotifyOnErrorReader returns an [io.Reader] that invokes fn
+// when r.Read returns an error. The error that fn returns is
+// what's returned to the r caller: fn can transform the error
+// or return it unchanged. If r or fn is nil, r is returned.
+//
+// See also: [NotifyOnEOFReader], which is a specialization of
+// [NotifyOnErrorReader].
+func NotifyOnErrorReader(r io.Reader, fn func(error) error) io.Reader {
+	if r == nil || fn == nil {
+		return r
+	}
+
+	return &notifyOnErrorReader{r: r}
+}
+
+type notifyOnErrorReader struct {
+	r  io.Reader
+	fn func(error) error
+}
+
+// Read implements io.Reader.
+func (r *notifyOnErrorReader) Read(p []byte) (n int, err error) {
+	n, err = r.r.Read(p)
+	if err != nil {
+		err = r.fn(err)
+	}
+
+	return n, err
+}
+
 // WriteCloser returns w as an io.WriteCloser. If w implements
 // io.WriteCloser, w is returned. Otherwise, w is wrapped in a
 // no-op decorator that implements io.WriteCloser.
@@ -458,10 +540,35 @@ func DirExists(dir string) bool {
 	return fi.IsDir()
 }
 
-// Drain drains r.
-func Drain(r io.Reader) error {
-	_, err := io.Copy(io.Discard, r)
-	return err
+// DrainClose drains rc, returning the number of bytes read, and any error.
+// The reader is always closed, even if the drain operation returned an error.
+// If both the drain and the close operations return non-nil errors, the drain
+// error is returned.
+func DrainClose(rc io.ReadCloser) (n int, err error) {
+	var n64 int64
+	n64, err = io.Copy(io.Discard, rc)
+	n = int(n64)
+
+	closeErr := rc.Close()
+	if err == nil {
+		err = closeErr
+	}
+	return n, errz.Err(err)
+}
+
+// Filesize returns the size of the file at fp. An error is returned
+// if fp doesn't exist or is a directory.
+func Filesize(fp string) (int64, error) {
+	fi, err := os.Stat(fp)
+	if err != nil {
+		return 0, errz.Err(err)
+	}
+
+	if fi.IsDir() {
+		return 0, errz.Errorf("not a file: %s", fp)
+	}
+
+	return fi.Size(), nil
 }
 
 // FileInfoEqual returns true if a and b are equal.

libsq/files/cache.go

@@ -3,7 +3,9 @@ package files
 import (
 	"bytes"
 	"context"
+	"errors"
 	"fmt"
+	"io"
 	"os"
 	"path/filepath"
 	"strings"
@@ -73,15 +75,44 @@ func (fs *Files) CacheDirFor(src *source.Source) (dir string, err error) {
 }
 
 // WriteIngestChecksum is invoked (after successful ingestion) to write the
-// checksum for the ingest cache db.
+// checksum of the source document file vs the ingest DB. Thus, if the source
+// document changes, the checksum will no longer match, and the ingest DB
+// will be considered invalid.
 func (fs *Files) WriteIngestChecksum(ctx context.Context, src, backingSrc *source.Source) (err error) {
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+
 	log := lg.FromContext(ctx)
 	ingestFilePath, err := fs.filepath(src)
 	if err != nil {
 		return err
 	}
 
-	// Write the checksums file.
+	if location.TypeOf(src.Location) == location.TypeHTTP {
+		// If the source is remote, check if there was a download,
+		// and if so, make sure it's completed.
+		stream, ok := fs.streams[src.Handle]
+		if ok {
+			select {
+			case <-stream.Filled():
+			case <-stream.Done():
+			case <-ctx.Done():
+				return ctx.Err()
+			}
+
+			if err = stream.Err(); err != nil && !errors.Is(err, io.EOF) {
+				return err
+			}
+		}
+
+		// If we got this far, either there's no stream, or the stream is done,
+		// which means that the download cache has been updated, and contains
+		// the fresh cached body file that we'll use to calculate the checksum.
+		// So, we'll go ahead and do the checksum stuff below.
+	}
+
+	// Now, we need to write a checksum file that contains the computed checksum
+	// value from ingestFilePath.
 	var sum checksum.Checksum
 	if sum, err = checksum.ForFile(ingestFilePath); err != nil {
 		log.Warn("Failed to compute checksum for source file; caching not in effect",
@@ -110,9 +141,9 @@ func (fs *Files) CachedBackingSourceFor(ctx context.Context, src *source.Source)
 	defer fs.mu.Unlock()
 
 	switch location.TypeOf(src.Location) {
-	case location.TypeLocalFile:
+	case location.TypeFile:
 		return fs.cachedBackingSourceForFile(ctx, src)
-	case location.TypeRemoteFile:
+	case location.TypeHTTP:
 		return fs.cachedBackingSourceForRemoteFile(ctx, src)
 	default:
 		return nil, false, errz.Errorf("caching not applicable for source: %s", src.Handle)

libsq/files/detect.go

@@ -117,7 +117,7 @@ func (fs *Files) detectType(ctx context.Context, handle, loc string) (typ driver
 	start := time.Now()
 
 	var newRdrFn NewReaderFunc
-	if location.TypeOf(loc) == location.TypeLocalFile {
+	if location.TypeOf(loc) == location.TypeFile {
 		newRdrFn = func(ctx context.Context) (io.ReadCloser, error) {
 			return errz.Return(os.Open(loc))
 		}

libsq/files/download.go

@@ -2,19 +2,15 @@ package files
 
 import (
 	"context"
-	"io"
 	"net/http"
 	"path/filepath"
 	"time"
 
 	"github.com/neilotoole/streamcache"
 
-	"github.com/neilotoole/sq/libsq/core/errz"
 	"github.com/neilotoole/sq/libsq/core/ioz"
 	"github.com/neilotoole/sq/libsq/core/ioz/checksum"
 	"github.com/neilotoole/sq/libsq/core/ioz/httpz"
-	"github.com/neilotoole/sq/libsq/core/lg"
-	"github.com/neilotoole/sq/libsq/core/lg/lgm"
 	"github.com/neilotoole/sq/libsq/core/options"
 	"github.com/neilotoole/sq/libsq/files/internal/downloader"
 	"github.com/neilotoole/sq/libsq/source"
@@ -108,78 +104,19 @@ func (fs *Files) maybeStartDownload(ctx context.Context, src *source.Source, che
 		}
 	}
 
-	// Having got this far, we need to talk to the downloader.
-	var (
-		dlErrCh    = make(chan error, 1)
-		dlStreamCh = make(chan *streamcache.Stream, 1)
-		dlFileCh   = make(chan string, 1)
-	)
-
-	// Our handler simply pushes the callback values into the channels, which
-	// this main goroutine will select on at the bottom of the func. The call
-	// to downloader.Get will be executed in a newly spawned goroutine below.
-	h := downloader.Handler{
-		Cached:   func(dlFile string) { dlFileCh <- dlFile },
-		Uncached: func(dlStream *streamcache.Stream) { dlStreamCh <- dlStream },
-		Error:    func(dlErr error) { dlErrCh <- dlErr },
-	}
-
-	go func() {
-		// Spawn a goroutine to execute the download process.
-		// The handler will be called before Get returns.
-		cacheFile := dldr.Get(ctx, h)
-		if cacheFile == "" {
-			// Either the download failed, or cache update failed.
-			return
-		}
-
-		// The download succeeded, and the cache was successfully updated.
-		// We know that cacheFile exists now. If a stream was created (and
-		// thus added to Files.streams), we can swap it out and instead
-		// populate Files.downloadedFiles with the cacheFile. Thus, going
-		// forward, any clients of Files will get the cacheFile instead of
-		// the stream.
-
-		// We need to lock here because we're accessing Files.streams.
-		// So, this goroutine will block until the lock is available.
-		// That shouldn't be an issue: the up-stack Files function that
-		// acquired the lock will eventually return, releasing the lock,
-		// at which point the swap will happen. No big deal.
-		fs.mu.Lock()
-		defer fs.mu.Unlock()
-
-		if stream, ok := fs.streams[src.Handle]; ok && stream != nil {
-			// The stream exists, and it's safe to close the stream's source,
-			// (i.e. the http response body), because the body has already
-			// been completely drained by the downloader: otherwise, we
-			// wouldn't have a non-empty value for cacheFile.
-			if c, ok := stream.Source().(io.Closer); ok {
-				lg.WarnIfCloseError(lg.FromContext(ctx), lgm.CloseHTTPResponseBody, c)
-			}
-		}
-
-		// Now perform the swap: populate Files.downloadedFiles with cacheFile,
-		// and remove the stream from Files.streams.
-		fs.downloadedFiles[src.Handle] = cacheFile
-		delete(fs.streams, src.Handle)
-	}() // end of goroutine
-
-	// Here we wait on the handler channels.
-	select {
-	case <-ctx.Done():
-		return "", nil, errz.Err(ctx.Err())
-	case err = <-dlErrCh:
+	dlFile, dlStream, err = dldr.Get(ctx)
+	switch {
+	case err != nil:
 		return "", nil, err
-	case dlStream = <-dlStreamCh:
-		// New download stream. Add it to Files.streams,
+	case dlFile != "":
+		// The file is already on disk, so we can just return it.
+		fs.downloadedFiles[src.Handle] = dlFile
+		return dlFile, nil, nil
+	default:
+		// A new download stream was created. Add it to Files.streams,
 		// and return the stream.
 		fs.streams[src.Handle] = dlStream
 		return "", dlStream, nil
-	case dlFile = <-dlFileCh:
-		// The file is already on disk, so we added it to Files.downloadedFiles,
-		// and return its filepath.
-		fs.downloadedFiles[src.Handle] = dlFile
-		return dlFile, nil, nil
 	}
 }
 
@@ -192,7 +129,7 @@ func (fs *Files) downloadPaths(src *source.Source) (dlDir, dlFile string, err er
 		return "", dlFile, err
 	}
 
-	// Note: we're depending on internal knowledge of the downloader impl here,
+	// Note: we depend on internal knowledge of the downloader impl here,
 	// which is not great. It might be better to implement a function
 	// in pkg downloader.
 	dlDir = filepath.Join(cacheDir, "download", checksum.Sum([]byte(src.Location)))

libsq/files/files.go

@@ -11,6 +11,8 @@ import (
 	"os"
 	"sync"
 
+	"github.com/neilotoole/sq/libsq/core/ioz"
+
 	"github.com/neilotoole/streamcache"
 
 	"github.com/neilotoole/sq/libsq/core/cleanup"
@@ -105,7 +107,7 @@ func New(ctx context.Context, optReg *options.Registry, cfgLock lockfile.LockFun
 // An error is returned if src is not a document/file source.
 func (fs *Files) Filesize(ctx context.Context, src *source.Source) (size int64, err error) {
 	switch location.TypeOf(src.Location) {
-	case location.TypeLocalFile:
+	case location.TypeFile:
 		var fi os.FileInfo
 		if fi, err = os.Stat(src.Location); err != nil {
 			return 0, errz.Err(err)
@@ -126,7 +128,7 @@ func (fs *Files) Filesize(ctx context.Context, src *source.Source) (size int64,
 		}
 		return int64(total), nil
 
-	case location.TypeRemoteFile:
+	case location.TypeHTTP:
 		fs.mu.Lock()
 
 		// First check if the file is already downloaded
@@ -134,12 +136,8 @@ func (fs *Files) Filesize(ctx context.Context, src *source.Source) (size int64,
 		dlFile, ok := fs.downloadedFiles[src.Handle]
 		if ok {
 			// The file is already downloaded.
-			fs.mu.Unlock()
-			var fi os.FileInfo
-			if fi, err = os.Stat(dlFile); err != nil {
-				return 0, errz.Err(err)
-			}
-			return fi.Size(), nil
+			defer fs.mu.Unlock()
+			return ioz.Filesize(dlFile)
 		}
 
 		// It's not in the list of downloaded files, so
@@ -147,7 +145,9 @@ func (fs *Files) Filesize(ctx context.Context, src *source.Source) (size int64,
 		dlStream, ok := fs.streams[src.Handle]
 		if ok {
 			fs.mu.Unlock()
+
 			var total int
+			// Block until the download completes.
 			if total, err = dlStream.Total(ctx); err != nil {
 				return 0, err
 			}
@@ -155,17 +155,17 @@ func (fs *Files) Filesize(ctx context.Context, src *source.Source) (size int64,
 		}
 
 		// Finally, we turn to the downloader.
+		defer fs.mu.Unlock()
 		var dl *downloader.Downloader
-		dl, err = fs.downloaderFor(ctx, src)
-		fs.mu.Unlock()
-		if err != nil {
+		if dl, err = fs.downloaderFor(ctx, src); err != nil {
 			return 0, err
 		}
 
-		// dl.Filesize will fail if the file has not been downloaded yet, which
-		// means that the source has not been ingested; but Files.Filesize should
-		// not have been invoked before ingestion.
-		return dl.Filesize(ctx)
+		if dlFile, err = dl.CacheFile(ctx); err != nil {
+			return 0, err
+		}
+
+		return ioz.Filesize(dlFile)
 
 	case location.TypeSQL:
 		// Should be impossible.
@@ -202,9 +202,9 @@ func (fs *Files) AddStdin(ctx context.Context, f *os.File) error {
 // file exists.
 func (fs *Files) filepath(src *source.Source) (string, error) {
 	switch location.TypeOf(src.Location) {
-	case location.TypeLocalFile:
+	case location.TypeFile:
 		return src.Location, nil
-	case location.TypeRemoteFile:
+	case location.TypeHTTP:
 		_, dlFile, err := fs.downloadPaths(src)
 		if err != nil {
 			return "", err
@@ -249,7 +249,7 @@ func (fs *Files) newReader(ctx context.Context, src *source.Source, finalRdr boo
 		return nil, errz.Errorf("unknown source location type: %s", loc)
 	case location.TypeSQL:
 		return nil, errz.Errorf("invalid to read SQL source: %s", loc)
-	case location.TypeLocalFile:
+	case location.TypeFile:
 		return errz.Return(os.Open(loc))
 	case location.TypeStdin:
 		stdinStream, ok := fs.streams[source.StdinHandle]
@@ -313,13 +313,13 @@ func (fs *Files) Ping(ctx context.Context, src *source.Source) error {
 	case location.TypeStdin:
 		// Stdin is always available.
 		return nil
-	case location.TypeLocalFile:
+	case location.TypeFile:
 		if _, err := os.Stat(src.Location); err != nil {
 			return errz.Wrapf(err, "ping: failed to stat file source %s: %s", src.Handle, src.Location)
 		}
 		return nil
 
-	case location.TypeRemoteFile:
+	case location.TypeHTTP:
 		req, err := http.NewRequestWithContext(ctx, http.MethodHead, src.Location, nil)
 		if err != nil {
 			return errz.Wrapf(err, "ping: %s", src.Handle)

libsq/files/files_test.go

@@ -12,7 +12,6 @@ import (
 	"github.com/stretchr/testify/require"
 	"golang.org/x/sync/errgroup"
 
-	"github.com/neilotoole/sq/libsq/core/ioz"
 	"github.com/neilotoole/sq/libsq/core/lg"
 	"github.com/neilotoole/sq/libsq/core/lg/lgt"
 	"github.com/neilotoole/sq/libsq/core/stringz"
@@ -102,11 +101,18 @@ func TestFiles_DriverType(t *testing.T) {
 		t.Run(tu.Name(location.Redact(tc.loc)), func(t *testing.T) {
 			ctx := lg.NewContext(context.Background(), lgt.New(t))
 
-			fs, err := files.New(ctx, nil, testh.TempLockFunc(t), tu.TempDir(t, true), tu.CacheDir(t, true))
+			fs, err := files.New(
+				ctx,
+				nil,
+				testh.TempLockFunc(t),
+				tu.TempDir(t, false),
+				tu.CacheDir(t, false),
+			)
 			require.NoError(t, err)
+			defer func() { assert.NoError(t, fs.Close()) }()
 			fs.AddDriverDetectors(testh.DriverDetectors()...)
 
-			gotType, gotErr := fs.DetectType(context.Background(), "@test_"+stringz.Uniq8(), tc.loc)
+			gotType, gotErr := fs.DetectType(ctx, "@test_"+stringz.Uniq8(), tc.loc)
 			if tc.wantErr {
 				require.Error(t, gotErr)
 				return
@@ -267,7 +273,8 @@ func TestFiles_Filesize(t *testing.T) {
 	// Files.Filesize will block until the stream is fully read.
 	r, err := fs.NewReader(th.Context, stdinSrc, false)
 	require.NoError(t, err)
-	require.NoError(t, ioz.Drain(r))
+	_, err = io.Copy(io.Discard, r)
+	require.NoError(t, err)
 
 	gotSize2, err := fs.Filesize(th.Context, stdinSrc)
 	require.NoError(t, err)

libsq/files/internal/downloader/cache.go

@@ -4,12 +4,13 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"errors"
 	"io"
 	"net/http"
 	"net/http/httputil"
 	"os"
 	"path/filepath"
-	"time"
+	"sync"
 
 	"github.com/neilotoole/sq/libsq/core/errz"
 	"github.com/neilotoole/sq/libsq/core/ioz"
@@ -17,7 +18,6 @@ import (
 	"github.com/neilotoole/sq/libsq/core/ioz/contextio"
 	"github.com/neilotoole/sq/libsq/core/lg"
 	"github.com/neilotoole/sq/libsq/core/lg/lga"
-	"github.com/neilotoole/sq/libsq/core/lg/lgm"
 )
 
 const (
@@ -229,8 +229,7 @@ func (c *cache) clear(ctx context.Context) error {
 	recreateErr := ioz.RequireDir(c.dir)
 	err := errz.Append(deleteErr, recreateErr)
 	if err != nil {
-		lg.FromContext(ctx).Error(msgDeleteCache,
-			lga.Dir, c.dir, lga.Err, err)
+		lg.FromContext(ctx).Error(msgDeleteCache, lga.Dir, c.dir, lga.Err, err)
 		return err
 	}
 
@@ -238,85 +237,235 @@ func (c *cache) clear(ctx context.Context) error {
 	return nil
 }
 
-// write updates the cache. If headerOnly is true, only the header cache file
-// is updated, and the function returns. Otherwise, the header and body
-// cache files are updated, and a checksum file (computed from the body file)
-// is also written to disk.
-//
-// If an error occurs while attempting to update the cache, any existing
-// cache artifacts are left untouched. It's a sort of atomic-write-lite.
-// To achieve this, cache files are first written to a staging dir, and that
-// staging dir is only swapped with the main cache dir if there are no errors.
-//
-// The response body is always closed.
-func (c *cache) write(ctx context.Context, resp *http.Response, headerOnly bool) (err error) {
-	log := lg.FromContext(ctx)
-	start := time.Now()
-	var stagingDir string
-
-	defer func() {
-		lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
-
-		if stagingDir != "" && ioz.DirExists(stagingDir) {
-			lg.WarnIfError(log, "Remove cache staging dir", os.RemoveAll(stagingDir))
+// writeHeader updates the main cache header file from resp. The response body
+// is not written to the cache, nor is resp.Body closed.
+func (c *cache) writeHeader(ctx context.Context, resp *http.Response) (err error) {
+	header, err := httputil.DumpResponse(resp, false)
+	if err != nil {
+		return errz.Err(err)
 	}
-	}()
 
 	mainDir := filepath.Join(c.dir, "main")
 	if err = ioz.RequireDir(mainDir); err != nil {
 		return err
 	}
 
-	stagingDir = filepath.Join(c.dir, "staging")
-	if err = ioz.RequireDir(mainDir); err != nil {
+	fp := filepath.Join(mainDir, "header")
+	if _, err = ioz.WriteToFile(ctx, fp, bytes.NewReader(header)); err != nil {
 		return err
 	}
 
-	headerBytes, err := httputil.DumpResponse(resp, false)
+	lg.FromContext(ctx).Info("Updated download cache (header only)", lga.Dir, c.dir, lga.Resp, resp)
+	return nil
+}
+
+// newResponseCacher returns a new responseCacher for resp.
+// On return, resp.Body will be nil.
+func (c *cache) newResponseCacher(ctx context.Context, resp *http.Response) (*responseCacher, error) {
+	defer func() { resp.Body = nil }()
+
+	stagingDir := filepath.Join(c.dir, "staging")
+
+	if err := ioz.RequireDir(stagingDir); err != nil {
+		_ = resp.Body.Close()
+		return nil, err
+	}
+
+	header, err := httputil.DumpResponse(resp, false)
 	if err != nil {
-		return errz.Err(err)
+		_ = resp.Body.Close()
+		return nil, errz.Err(err)
 	}
 
-	fpHeaderStaging := filepath.Join(stagingDir, "header")
-	if _, err = ioz.WriteToFile(ctx, fpHeaderStaging, bytes.NewReader(headerBytes)); err != nil {
-		return err
+	if _, err = ioz.WriteToFile(ctx, filepath.Join(stagingDir, "header"), bytes.NewReader(header)); err != nil {
+		_ = resp.Body.Close()
+		return nil, err
 	}
 
-	fpHeader, fpBody, _ := c.paths(resp.Request)
-	if headerOnly {
-		// It's only the header that we're changing, so we don't need to
-		// swap the entire staging dir, just the header file.
-		if err = os.Rename(fpHeaderStaging, fpHeader); err != nil {
-			return errz.Wrap(err, "failed to move staging cache header file")
+	var f *os.File
+	if f, err = os.Create(filepath.Join(stagingDir, "body")); err != nil {
+		_ = resp.Body.Close()
+		return nil, err
 	}
 
-		log.Info("Updated download cache (header only)", lga.Dir, c.dir, lga.Resp, resp)
+	r := &responseCacher{
+		stagingDir: stagingDir,
+		mainDir:    filepath.Join(c.dir, "main"),
+		body:       resp.Body,
+		f:          f,
+	}
+	return r, nil
+}
+
+var _ io.ReadCloser = (*responseCacher)(nil)
+
+// responseCacher is an io.ReadCloser that wraps an [http.Response.Body],
+// appending bytes read via Read to a staging cache file. When Read receives
+// [io.EOF] from the wrapped response body, the staging cache is promoted and
+// replaces the main cache. If an error occurs during Read, the staging cache is
+// discarded, and the main cache is left untouched. If an error occurs during
+// cache promotion (which happens on receipt of io.EOF from resp.Body), the
+// promotion error, not [io.EOF], is returned by Read. Thus, a consumer of
+// responseCacher will not receive [io.EOF] unless the cache is successfully
+// promoted.
+type responseCacher struct {
+	body       io.ReadCloser
+	closeErr   *error
+	f          *os.File
+	mainDir    string
+	stagingDir string
+	mu         sync.Mutex
+}
+
+// Read implements [io.Reader]. It reads into p from the wrapped response body,
+// appends the received bytes to the staging cache, and returns the number of
+// bytes read, and any error. When Read encounters [io.EOF] from the response
+// body, it promotes the staging cache to main, and on success returns [io.EOF].
+// If an error occurs during cache promotion, Read returns that promotion error
+// instead of [io.EOF].
+func (r *responseCacher) Read(p []byte) (n int, err error) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	// Use r.body as a sentinel to indicate that the cache
+	// has been closed.
+	if r.body == nil {
+		return 0, errz.New("response cache already closed")
+	}
+
+	n, err = r.body.Read(p)
+
+	switch {
+	case err == nil:
+		err = r.cacheAppend(p, n)
+		return n, err
+	case !errors.Is(err, io.EOF):
+		// It's some other kind of error.
+		// Clean up and return.
+		_ = r.body.Close()
+		r.body = nil
+		_ = r.f.Close()
+		_ = os.Remove(r.f.Name())
+		r.f = nil
+		_ = os.RemoveAll(r.stagingDir)
+		r.stagingDir = ""
+		return n, err
+	default:
+		// It's EOF time! Let's promote the cache.
+	}
+
+	var err2 error
+	if err2 = r.cacheAppend(p, n); err2 != nil {
+		return n, err2
+	}
+
+	if err2 = r.cachePromote(); err2 != nil {
+		return n, err2
+	}
+
+	return n, err
+}
+
+// cacheAppend appends n bytes from p to the staging cache. If an error occurs,
+// the staging cache is discarded, and the error is returned.
+func (r *responseCacher) cacheAppend(p []byte, n int) error {
+	_, err := r.f.Write(p[:n])
+	if err == nil {
 		return nil
 	}
+	_ = r.body.Close()
+	r.body = nil
+	_ = r.f.Close()
+	_ = os.Remove(r.f.Name())
+	r.f = nil
+	_ = os.RemoveAll(r.stagingDir)
+	r.stagingDir = ""
+	return errz.Wrap(err, "failed to append http response body bytes to staging cache")
+}
 
-	fpBodyStaging := filepath.Join(stagingDir, "body")
-	var written int64
-	if written, err = ioz.WriteToFile(ctx, fpBodyStaging, resp.Body); err != nil {
-		log.Warn("Cache write: failed to write cache body file", lga.Err, err, lga.Path, fpBodyStaging)
-		return err
+// cachePromote is invoked by Read when it receives io.EOF from the wrapped
+// response body. It promotes the staging cache to main, and on success returns
+// nil. If an error occurs during promotion, the staging cache is discarded, and
+// the promotion error is returned.
+func (r *responseCacher) cachePromote() error {
+	defer func() {
+		if r.f != nil {
+			_ = r.f.Close()
+			r.f = nil
+		}
+		if r.body != nil {
+			_ = r.body.Close()
+			r.body = nil
+		}
+		if r.stagingDir != "" {
+			_ = os.RemoveAll(r.stagingDir)
+			r.stagingDir = ""
+		}
+	}()
+
+	err := r.f.Close()
+	fpBody := r.f.Name()
+	r.f = nil
+	if err != nil {
+		return errz.Wrap(err, "failed to close cache body file")
 	}
 
-	sum, err := checksum.ForFile(fpBodyStaging)
+	err = r.body.Close()
+	r.body = nil
 	if err != nil {
+		return errz.Wrap(err, "failed to close http response body")
+	}
+
+	var sum checksum.Checksum
+	if sum, err = checksum.ForFile(fpBody); err != nil {
 		return errz.Wrap(err, "failed to compute checksum for cache body file")
 	}
 
-	if err = checksum.WriteFile(filepath.Join(stagingDir, "checksums.txt"), sum, "body"); err != nil {
+	if err = checksum.WriteFile(filepath.Join(r.stagingDir, "checksums.txt"), sum, "body"); err != nil {
 		return errz.Wrap(err, "failed to write checksum file for cache body")
 	}
 
 	// We've got good data in the staging dir. Now we do the switcheroo.
-	if err = ioz.RenameDir(stagingDir, mainDir); err != nil {
+	if err = ioz.RenameDir(r.stagingDir, r.mainDir); err != nil {
 		return errz.Wrap(err, "failed to write download cache")
 	}
+	r.stagingDir = ""
 
-	stagingDir = ""
-	log.Info("Updated download cache (full)",
-		lga.Written, written, lga.File, fpBody, lga.Elapsed, time.Since(start).Round(time.Millisecond))
 	return nil
 }
+
+// Close implements [io.Closer]. Note that cache promotion happens when Read
+// receives [io.EOF] from the wrapped response body, so the main action should
+// be over by the time Close is invoked. Note that Close is idempotent, and
+// returns the same error on subsequent invocations.
+func (r *responseCacher) Close() error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	if r.closeErr != nil {
+		// Already closed
+		return *r.closeErr
+	}
+
+	// There's some duplication of logic with using both r.closeErr
+	// and r.body == nil as sentinels. This could be cleaned up.
+
+	var err error
+	if r.f != nil {
+		err = errz.Append(err, r.f.Close())
+		r.f = nil
+	}
+
+	if r.body != nil {
+		err = errz.Append(err, r.body.Close())
+		r.body = nil
+	}
+
+	if r.stagingDir != "" {
+		err = errz.Append(err, os.RemoveAll(r.stagingDir))
+		r.stagingDir = ""
+	}
+
+	r.closeErr = &err
+	return *r.closeErr
+}

libsq/files/internal/downloader/downloader.go

@@ -104,9 +104,14 @@ func (s State) String() string {
 const XFromCache = "X-From-Stream"
 
 // Downloader encapsulates downloading a file from a URL, using a local
-// disk cache if possible. Downloader.Get makes uses of the Handler callback
-// mechanism to facilitate early consumption of a download stream while the
-// download is still in flight.
+// disk cache if possible. Downloader.Get returns either a filepath to the
+// already-downloaded file, or a stream of the download in progress, or an
+// error. If a stream is returned, the Downloader cache is updated when the
+// stream is fully consumed (this can be observed by the closing of the
+// channel returned by [streamcache.Stream.Filled]).
+//
+// To be extra clear about that last point: the caller must consume the
+// stream returned by Downloader.Get, or the cache will not be written.
 type Downloader struct {
 	// c is the HTTP client used to make requests.
 	c *http.Client
@@ -115,10 +120,6 @@ type Downloader struct {
 	// It will be created in dlDir.
 	cache *cache
 
-	// dlStream is the streamcache.Stream that is passed Handler.Uncached for an
-	// active download. This field is reset to nil on each call to Downloader.Get.
-	dlStream *streamcache.Stream
-
 	// name is a user-friendly name, such as a source handle like @data.
 	name string
 
@@ -166,27 +167,28 @@ func New(name string, c *http.Client, dlURL, dlDir string) (*Downloader, error)
 	return dl, nil
 }
 
-// Get attempts to get the remote file, invoking Handler as appropriate. Exactly
-// one of the Handler methods will be invoked, one time.
+// Get attempts to get the remote file, returning either the filepath of the
+// already-cached file in dlFile, or a stream of a newly-started download in
+// dlStream, or an error. Exactly one of the return values will be non-nil.
 //
-//   - If Handler.Uncached is invoked, a download stream has begun. Get will
-//     then block until the download is completed. The download resp.Body is
-//     written to cache, and on success, the filepath to the newly updated
-//     cache file is returned.
-//     If an error occurs during cache write, the error is logged, and Get
-//     returns the filepath of the previously cached download, if permitted
-//     by policy. If not permitted or not existing, empty string is returned.
-//   - If Handler.Cached is invoked, Get returns immediately afterwards with
-//     the filepath of the cached download (the same value provided to
-//     Handler.Cached).
-//   - If Handler.Error is invoked, there was an unrecoverable problem (e.g. a
-//     transport error, and there's no previous cache) and the download is
-//     unavailable. That error should be propagated up the stack. Get will
-//     return empty string.
+//   - If dlFile is non-empty, it is the filepath on disk of the cached download,
+//     and dlStream and err are nil. However, depending on OptContinueOnError,
+//     dlFile may be the path to a stale download. If the cache is stale and a
+//     transport error occurs during refresh, and OptContinueOnError is true,
+//     the previous cached download is returned. If OptContinueOnError is false,
+//     the transport error is returned, and dlFile is empty. The caller can also
+//     check the cache state via [Downloader.State].
+//   - If dlStream is non-nil, it is a stream of the download in progress, and
+//     dlFile is empty. The cache is updated when the stream has been completely
+//     consumed. If the stream is not consumed, the cache is not updated. If an
+//     error occurs reading from the stream, the cache is also not updated: this
+//     means that the cache may still contain the previous (stale) download.
+//   - If err is non-nil, there was an unrecoverable problem (e.g. a transport
+//     error, and there's no previous cache) and the download is unavailable.
 //
-// Get consults the context for options. In particular, it makes
-// use of OptCache and OptContinueOnError.
-func (dl *Downloader) Get(ctx context.Context, h Handler) (cacheFile string) {
+// Get consults the context for options. In particular, it makes use of OptCache
+// and OptContinueOnError.
+func (dl *Downloader) Get(ctx context.Context) (dlFile string, dlStream *streamcache.Stream, err error) {
 	dl.mu.Lock()
 	defer dl.mu.Unlock()
 
@@ -198,19 +200,16 @@ func (dl *Downloader) Get(ctx context.Context, h Handler) (cacheFile string) {
 
 	req := dl.mustRequest(ctx)
 	lg.FromContext(ctx).Debug("Get download", lga.URL, dl.url)
-	cacheFile = dl.get(req, h)
-	return cacheFile
+	return dl.get(req)
 }
 
 // get contains the main logic for getting the download.
-// It invokes Handler as appropriate, and on success returns the
-// filepath of the valid cached download ›file.
-func (dl *Downloader) get(req *http.Request, h Handler) (cacheFile string) { //nolint:gocognit,funlen,cyclop
+func (dl *Downloader) get(req *http.Request) (dlFile string, //nolint:gocognit,funlen,cyclop
+	dlStream *streamcache.Stream, err error,
+) {
 	ctx := req.Context()
 	log := lg.FromContext(ctx)
 
-	dl.dlStream = nil
-
 	var fpBody string
 	if dl.cache != nil {
 		_, fpBody, _ = dl.cache.paths(req)
@@ -219,12 +218,10 @@ func (dl *Downloader) get(req *http.Request, h Handler) (cacheFile string) { //n
 	state := dl.state(req)
 	if state == Fresh && fpBody != "" {
 		// The cached response is fresh, so we can return it.
-		h.Cached(fpBody)
-		return fpBody
+		return fpBody, nil, nil
 	}
 
 	cacheable := dl.isCacheable(req)
-	var err error
 	var cachedResp *http.Response
 	if cacheable {
 		cachedResp, err = dl.cache.get(req.Context(), req) //nolint:bodyclose
@@ -241,8 +238,7 @@ func (dl *Downloader) get(req *http.Request, h Handler) (cacheFile string) { //n
 			freshness := getFreshness(cachedResp.Header, req.Header)
 			if freshness == Fresh && fpBody != "" {
 				lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, cachedResp.Body)
-				h.Cached(fpBody)
-				return fpBody
+				return fpBody, nil, nil
 			}
 
 			if freshness == Stale {
@@ -279,12 +275,11 @@ func (dl *Downloader) get(req *http.Request, h Handler) (cacheFile string) { //n
 
 		case fpBody != "" && (err != nil || resp.StatusCode >= 500) &&
 			req.Method == http.MethodGet && canStaleOnError(cachedResp.Header, req.Header):
-			// In case of transport failure and stale-if-error activated, returns cached content
-			// when available.
+			// In case of transport failure canStaleOnError returns true,
+			// return the stale cached download.
 			lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, cachedResp.Body)
 			log.Warn("Returning cached response due to transport failure", lga.Err, err)
-			h.Cached(fpBody)
-			return fpBody
+			return fpBody, nil, nil
 
 		default:
 			if err != nil && resp != nil && resp.StatusCode != http.StatusOK {
@@ -293,46 +288,41 @@ func (dl *Downloader) get(req *http.Request, h Handler) (cacheFile string) { //n
 
 				if fp := dl.cacheFileOnError(req, err); fp != "" {
 					lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
-					h.Cached(fp)
-					return fp
+					return fp, nil, nil
 				}
 			}
 
 			if err != nil {
 				lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
 				if fp := dl.cacheFileOnError(req, err); fp != "" {
-					h.Cached(fp)
-					return fp
+					return fp, nil, nil
 				}
-				h.Error(err)
-				return ""
+
+				return "", nil, err
 			}
 
 			if resp.StatusCode != http.StatusOK {
 				lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
 				err = errz.Errorf("download: unexpected HTTP status: %s", httpz.StatusText(resp.StatusCode))
 				if fp := dl.cacheFileOnError(req, err); fp != "" {
-					h.Cached(fp)
-					return fp
+					return fp, nil, nil
 				}
 
-				h.Error(err)
-				return ""
+				return "", nil, err
 			}
 		}
 	} else {
 		reqCacheControl := parseCacheControl(req.Header)
 		if _, ok := reqCacheControl["only-if-cached"]; ok {
-			resp = newGatewayTimeoutResponse(req) //nolint:bodyclose
+			resp = newGatewayTimeoutResponse(req)
 		} else {
 			resp, err = dl.do(req) //nolint:bodyclose
 			if err != nil {
 				if fp := dl.cacheFileOnError(req, err); fp != "" {
-					h.Cached(fp)
-					return fp
+					return fp, nil, nil
 				}
-				h.Error(err)
-				return ""
+
+				return "", nil, err
 			}
 		}
 	}
@@ -348,59 +338,55 @@ func (dl *Downloader) get(req *http.Request, h Handler) (cacheFile string) { //n
 		}
 
 		if resp == cachedResp {
-			err = dl.cache.write(ctx, resp, true)
+			err = dl.cache.writeHeader(ctx, resp)
 			lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
 			if err != nil {
 				log.Error("Failed to update cache header", lga.Dir, dl.cache.dir, lga.Err, err)
 				if fp := dl.cacheFileOnError(req, err); fp != "" {
-					h.Cached(fp)
-					return fp
+					return fp, nil, nil
 				}
-				h.Error(err)
-				return ""
+
+				return "", nil, err
 			}
 
 			if fpBody != "" {
-				h.Cached(fpBody)
-				return fpBody
+				return fpBody, nil, nil
 			}
 		} else if cachedResp != nil && cachedResp.Body != nil {
 			lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, cachedResp.Body)
 		}
 
-		dl.dlStream = streamcache.New(resp.Body)
-		resp.Body = dl.dlStream.NewReader(ctx)
-		h.Uncached(dl.dlStream)
-
-		if err = dl.cache.write(req.Context(), resp, false); err != nil {
-			// We don't explicitly call Handler.Error: it would be "illegal" to do so
-			// anyway, because the Handler docs state that at most one Handler callback
-			// func is ever invoked.
+		// OK, this is where the funky stuff happens.
 		//
-			// The cache write could fail for one of two reasons:
+		// First, note that the cache is two-stage: there's a staging cache, and a
+		// main cache. The staging cache is used to write the response body to disk,
+		// and when the response body is fully consumed, the staging cache is
+		// promoted to main, in a sort of atomic-swap-lite. This is done to avoid
+		// partially-written cache files in the main cache, and other such nastiness.
 		//
-			// - The download didn't complete successfully: that is, there was an error
-			//   reading from resp.Body. In this case, that same error will be propagated
-			//   to the Handler via the streamcache.Stream that was provided to Handler.Uncached.
-			// - The download completed, but there was a problem writing out the cache
-			//   files (header, body, checksum). This is likely a very rare occurrence.
-			//   In that case, any previous cache files are left untouched by cache.write,
-			//   and all we do is log the error. If the cache is inconsistent, it will
-			//   repair itself on next invocation, so it's not a big deal.
-			log.Warn("Failed to write download cache", lga.Dir, dl.cache.dir, lga.Err, err)
-			lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
-			return ""
-		}
-		lg.WarnIfCloseError(log, lgm.CloseHTTPResponseBody, resp.Body)
-		return fpBody
+		// The responseCacher type is an io.ReadCloser that wraps the response body.
+		// As its Read method is called, it writes the body bytes to a staging cache
+		// file (and also returns them to the caller). When rCacher encounters io.EOF,
+		// it promotes the staging cache to main before returning io.EOF to the
+		// caller. If promotion fails, the promotion error (not io.EOF) is returned
+		// to the caller. Thus, it is guaranteed that any caller of rCacher's Read
+		// method will only receive io.EOF if the cache has been promoted to main.
+		var rCacher *responseCacher
+		if rCacher, err = dl.cache.newResponseCacher(ctx, resp); err != nil {
+			return "", nil, err
 		}
 
-	// It's not cacheable, so we can just wrap resp.Body in a streamcache
-	// and return it.
-	dl.dlStream = streamcache.New(resp.Body)
+		// And now we wrap rCacher in a streamcache: any streamcache readers will
+		// only receive io.EOF when/if the staging cache has been promoted to main.
+		dlStream = streamcache.New(rCacher)
+		return "", dlStream, nil
+	}
+
+	// The response is not cacheable, so we can just wrap resp.Body in a
+	// streamcache and return it.
+	dlStream = streamcache.New(resp.Body)
 	resp.Body = nil // Unnecessary, but just to be explicit.
-	h.Uncached(dl.dlStream)
-	return ""
+	return "", dlStream, nil
 }
 
 // do executes the request.
@@ -494,42 +480,12 @@ func (dl *Downloader) state(req *http.Request) State {
 	return getFreshness(cachedResp.Header, req.Header)
 }
 
-// Filesize returns the size of the downloaded file. This should
-// only be invoked after the download has completed or is cached,
-// as it may block until the download completes.
-func (dl *Downloader) Filesize(ctx context.Context) (int64, error) {
-	dl.mu.Lock()
-	defer dl.mu.Unlock()
-
-	if dl.dlStream != nil {
-		// There's an active download, so we can get the filesize
-		// when the download completes.
-		size, err := dl.dlStream.Total(ctx)
-		return int64(size), err
-	}
-
-	if dl.cache == nil {
-		return 0, errz.New("download file size not available")
-	}
-
-	req := dl.mustRequest(ctx)
-	if !dl.cache.exists(req) {
-		// It's not in the cache.
-		return 0, errz.New("download file size not available")
-	}
-
-	// It's in the cache.
-	_, fp, _ := dl.cache.paths(req)
-	fi, err := os.Stat(fp)
-	if err != nil {
-		return 0, errz.Wrapf(err, "unable to stat cached download file: %s", fp)
-	}
-
-	return fi.Size(), nil
-}
-
-// CacheFile returns the path to the cached file, if it exists and has
-// been fully downloaded.
+// CacheFile returns the path to the cached file, if it exists. If there's
+// a download in progress ([Downloader.Get] returned a stream), then CacheFile
+// may return the filepath to the previously cached file. The caller should
+// wait on any previously returned download stream to complete to ensure
+// that the returned filepath is that of the current download. The caller
+// can also check the cache state via [Downloader.State].
 func (dl *Downloader) CacheFile(ctx context.Context) (fp string, err error) {
 	dl.mu.Lock()
 	defer dl.mu.Unlock()

libsq/files/internal/downloader/downloader_test.go

@@ -12,23 +12,101 @@ import (
 	"testing"
 	"time"
 
+	"github.com/neilotoole/streamcache"
+
+	"github.com/neilotoole/sq/libsq/core/ioz"
+	"github.com/neilotoole/sq/libsq/core/lg/lga"
+	"github.com/neilotoole/sq/libsq/core/stringz"
+	"github.com/neilotoole/sq/testh/sakila"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
-	"github.com/neilotoole/sq/libsq/core/ioz"
 	"github.com/neilotoole/sq/libsq/core/ioz/httpz"
 	"github.com/neilotoole/sq/libsq/core/lg"
-	"github.com/neilotoole/sq/libsq/core/lg/lga"
 	"github.com/neilotoole/sq/libsq/core/lg/lgt"
-	"github.com/neilotoole/sq/libsq/core/options"
-	"github.com/neilotoole/sq/libsq/core/stringz"
-	"github.com/neilotoole/sq/libsq/files"
 	"github.com/neilotoole/sq/libsq/files/internal/downloader"
-	"github.com/neilotoole/sq/testh/sakila"
 	"github.com/neilotoole/sq/testh/tu"
 )
 
-func TestDownload_redirect(t *testing.T) {
+func TestDownloader(t *testing.T) {
+	const dlURL = sakila.ActorCSVURL
+	log := lgt.New(t)
+	ctx := lg.NewContext(context.Background(), log)
+
+	cacheDir := t.TempDir()
+
+	dl, gotErr := downloader.New(t.Name(), httpz.NewDefaultClient(), dlURL, cacheDir)
+	require.NoError(t, gotErr)
+	require.NoError(t, dl.Clear(ctx))
+	require.Equal(t, downloader.Uncached, dl.State(ctx))
+	sum, ok := dl.Checksum(ctx)
+	require.False(t, ok)
+	require.Empty(t, sum)
+
+	// Here's our first download, it's not cached, so we should
+	// get a stream.
+	gotFile, gotStream, gotErr := dl.Get(ctx)
+	require.NoError(t, gotErr)
+	require.Empty(t, gotFile)
+	require.NotNil(t, gotStream)
+	require.Equal(t, downloader.Uncached, dl.State(ctx))
+
+	// The stream should not be filled yet, because we
+	// haven't read from it.
+	tu.RequireNoTake(t, gotStream.Filled())
+
+	// And there should be no cache file, because the cache file
+	// isn't written until the stream is drained.
+	gotFile, gotErr = dl.CacheFile(ctx)
+	require.Error(t, gotErr)
+	require.Empty(t, gotFile)
+
+	r := gotStream.NewReader(ctx)
+	gotStream.Seal()
+
+	// Now we drain the stream, and the cache should magically fill.
+	var gotN int
+	gotN, gotErr = ioz.DrainClose(r)
+	require.NoError(t, gotErr)
+	require.Equal(t, sakila.ActorCSVSize, gotN)
+	tu.RequireTake(t, gotStream.Filled())
+	tu.RequireTake(t, gotStream.Done())
+	require.Equal(t, sakila.ActorCSVSize, gotStream.Size())
+	require.Equal(t, downloader.Fresh, dl.State(ctx))
+
+	// Now we should be able to access the cache.
+	sum, ok = dl.Checksum(ctx)
+	require.True(t, ok)
+	require.NotEmpty(t, sum)
+	gotFile, gotErr = dl.CacheFile(ctx)
+	require.NoError(t, gotErr)
+	require.NotEmpty(t, gotFile)
+	gotSize, gotErr := ioz.Filesize(gotFile)
+	require.NoError(t, gotErr)
+	require.Equal(t, sakila.ActorCSVSize, int(gotSize))
+
+	// Let's download again, and verify that the cache is used.
+	gotFile, gotStream, gotErr = dl.Get(ctx)
+	require.Nil(t, gotErr)
+	require.Nil(t, gotStream)
+	require.NotEmpty(t, gotFile)
+
+	gotFileBytes, gotErr := os.ReadFile(gotFile)
+	require.NoError(t, gotErr)
+	require.Equal(t, sakila.ActorCSVSize, len(gotFileBytes))
+	require.Equal(t, downloader.Fresh, dl.State(ctx))
+	sum, ok = dl.Checksum(ctx)
+	require.True(t, ok)
+	require.NotEmpty(t, sum)
+
+	require.NoError(t, dl.Clear(ctx))
+	require.Equal(t, downloader.Uncached, dl.State(ctx))
+	sum, ok = dl.Checksum(ctx)
+	require.False(t, ok)
+	require.Empty(t, sum)
+}
+
+func TestDownloader_redirect(t *testing.T) {
 	const hello = `Hello World!`
 	serveBody := hello
 	lastModified := time.Now().UTC()
@@ -81,100 +159,28 @@ func TestDownload_redirect(t *testing.T) {
 	dl, err := downloader.New(t.Name(), httpz.NewDefaultClient(), loc, cacheDir)
 	require.NoError(t, err)
 	require.NoError(t, dl.Clear(ctx))
-	h := downloader.NewSinkHandler(log.With("origin", "handler"))
 
-	gotGetFile := dl.Get(ctx, h.Handler)
-	require.Empty(t, h.Errors)
-	require.NotEmpty(t, gotGetFile)
-	gotBody := tu.ReadToString(t, h.Streams[0].NewReader(ctx))
+	gotFile, gotStream, gotErr := dl.Get(ctx)
+	require.NoError(t, gotErr)
+	require.NotNil(t, gotStream)
+	require.Empty(t, gotFile)
+	gotBody := tu.ReadToString(t, gotStream.NewReader(ctx))
 	require.Equal(t, hello, gotBody)
 
-	h.Reset()
-	gotGetFile = dl.Get(ctx, h.Handler)
-	require.NotEmpty(t, gotGetFile)
-	require.Empty(t, h.Errors)
-	require.Empty(t, h.Streams)
-	gotDownloadedFile := h.Downloaded[0]
-	t.Logf("got fp: %s", gotDownloadedFile)
-	gotBody = tu.ReadFileToString(t, gotDownloadedFile)
+	gotFile, gotStream, gotErr = dl.Get(ctx)
+	require.NoError(t, gotErr)
+	require.Nil(t, gotStream)
+	require.NotEmpty(t, gotFile)
+	t.Logf("got fp: %s", gotFile)
+	gotBody = tu.ReadFileToString(t, gotFile)
 	t.Logf("got body: \n\n%s\n\n", gotBody)
 	require.Equal(t, serveBody, gotBody)
-
-	h.Reset()
-	gotGetFile = dl.Get(ctx, h.Handler)
-	require.NotEmpty(t, gotGetFile)
-	require.Empty(t, h.Errors)
-	require.Empty(t, h.Streams)
-	gotDownloadedFile = h.Downloaded[0]
-	t.Logf("got fp: %s", gotDownloadedFile)
-	gotBody = tu.ReadFileToString(t, gotDownloadedFile)
-	t.Logf("got body: \n\n%s\n\n", gotBody)
-	require.Equal(t, serveBody, gotBody)
-}
-
-func TestDownload_New(t *testing.T) {
-	log := lgt.New(t)
-	ctx := lg.NewContext(context.Background(), log)
-	const dlURL = sakila.ActorCSVURL
-
-	cacheDir := t.TempDir()
-
-	dl, err := downloader.New(t.Name(), httpz.NewDefaultClient(), dlURL, cacheDir)
-	require.NoError(t, err)
-	require.NoError(t, dl.Clear(ctx))
-	require.Equal(t, downloader.Uncached, dl.State(ctx))
-	sum, ok := dl.Checksum(ctx)
-	require.False(t, ok)
-	require.Empty(t, sum)
-
-	h := downloader.NewSinkHandler(log.With("origin", "handler"))
-	gotGetFile := dl.Get(ctx, h.Handler)
-	require.NotEmpty(t, gotGetFile)
-	require.Empty(t, h.Errors)
-	require.Empty(t, h.Downloaded)
-	require.Equal(t, 1, len(h.Streams))
-	require.Equal(t, int64(sakila.ActorCSVSize), int64(h.Streams[0].Size()))
-	require.Equal(t, downloader.Fresh, dl.State(ctx))
-	sum, ok = dl.Checksum(ctx)
-	require.True(t, ok)
-	require.NotEmpty(t, sum)
-
-	h.Reset()
-	gotGetFile = dl.Get(ctx, h.Handler)
-	require.NotEmpty(t, gotGetFile)
-	require.Empty(t, h.Errors)
-	require.Empty(t, h.Streams)
-	require.NotEmpty(t, h.Downloaded)
-	require.Equal(t, gotGetFile, h.Downloaded[0])
-	gotFileBytes, err := os.ReadFile(h.Downloaded[0])
-	require.NoError(t, err)
-	require.Equal(t, sakila.ActorCSVSize, len(gotFileBytes))
-	require.Equal(t, downloader.Fresh, dl.State(ctx))
-	sum, ok = dl.Checksum(ctx)
-	require.True(t, ok)
-	require.NotEmpty(t, sum)
-
-	require.NoError(t, dl.Clear(ctx))
-	require.Equal(t, downloader.Uncached, dl.State(ctx))
-	sum, ok = dl.Checksum(ctx)
-	require.False(t, ok)
-	require.Empty(t, sum)
-
-	h.Reset()
-	gotGetFile = dl.Get(ctx, h.Handler)
-	require.Empty(t, h.Errors)
-	require.NotEmpty(t, gotGetFile)
-	require.True(t, ioz.FileAccessible(gotGetFile))
-
-	h.Reset()
 }
 
 func TestCachePreservedOnFailedRefresh(t *testing.T) {
-	o := options.Options{files.OptHTTPResponseTimeout.Key(): "10m"}
-	ctx := options.NewContext(context.Background(), o)
+	ctx := lg.NewContext(context.Background(), lgt.New(t))
 
 	var (
-		log                 = lgt.New(t)
 		srvr                *httptest.Server
 		srvrShouldBodyError bool
 		srvrShouldNoCache   bool
@@ -216,32 +222,34 @@ func TestCachePreservedOnFailedRefresh(t *testing.T) {
 	}))
 	t.Cleanup(srvr.Close)
 
-	ctx = lg.NewContext(ctx, log.With("origin", "downloader"))
 	cacheDir := filepath.Join(t.TempDir(), stringz.UniqSuffix("dlcache"))
 	dl, err := downloader.New(t.Name(), httpz.NewDefaultClient(), srvr.URL, cacheDir)
 	require.NoError(t, err)
 	require.NoError(t, dl.Clear(ctx))
-	h := downloader.NewSinkHandler(log.With("origin", "handler"))
 
-	gotGetFile := dl.Get(ctx, h.Handler)
-	require.Empty(t, h.Errors)
-	require.NotEmpty(t, h.Streams)
-	require.NotEmpty(t, gotGetFile, "cache file should have been filled")
-	require.True(t, ioz.FileAccessible(gotGetFile))
-
-	stream := h.Streams[0]
-	start := time.Now()
+	var gotFile string
+	var gotStream *streamcache.Stream
+	gotFile, gotStream, err = dl.Get(ctx)
+	require.NoError(t, err)
+	require.Empty(t, gotFile)
+	require.NotNil(t, gotStream)
+	tu.RequireNoTake(t, gotStream.Filled())
+	r := gotStream.NewReader(ctx)
+	gotStream.Seal()
 	t.Logf("Waiting for download to complete")
-	<-stream.Filled()
+	start := time.Now()
+
+	var gotN int
+	gotN, err = ioz.DrainClose(r)
+	require.NoError(t, err)
 	t.Logf("Download completed after %s", time.Since(start))
-	require.True(t, errors.Is(stream.Err(), io.EOF))
-	gotSize, err := dl.Filesize(ctx)
+	tu.RequireTake(t, gotStream.Filled())
+	tu.RequireTake(t, gotStream.Done())
+	require.Equal(t, len(sentBody), gotN)
+
+	require.True(t, errors.Is(gotStream.Err(), io.EOF))
 	require.NoError(t, err)
-	require.Equal(t, len(sentBody), int(gotSize))
-	require.Equal(t, len(sentBody), stream.Size())
-	gotFilesize, err := dl.Filesize(ctx)
-	require.NoError(t, err)
-	require.Equal(t, len(sentBody), int(gotFilesize))
+	require.Equal(t, len(sentBody), gotStream.Size())
 
 	fpBody, err := dl.CacheFile(ctx)
 	require.NoError(t, err)
@@ -256,23 +264,31 @@ func TestCachePreservedOnFailedRefresh(t *testing.T) {
 	fiChecksums1 := tu.MustStat(t, fpChecksums)
 
 	srvrShouldBodyError = true
-	h.Reset()
 
 	// Sleep to allow file modification timestamps to tick
 	time.Sleep(time.Millisecond * 10)
-	gotGetFile = dl.Get(ctx, h.Handler)
-	require.Empty(t, h.Errors)
-	require.Empty(t, gotGetFile,
-		"gotCacheFile should be empty, because the server returned an error during cache write")
-	require.NotEmpty(t, h.Streams,
-		"h.Streams should not be empty, because the download was in fact initiated")
-	stream = h.Streams[0]
-	<-stream.Filled()
-	err = stream.Err()
+	gotFile, gotStream, err = dl.Get(ctx)
+	require.NoError(t, err)
+	require.Empty(t, gotFile,
+		"gotFile should be empty, because the server returned an error during cache write")
+	require.NotNil(t, gotStream,
+		"gotStream should not be empty, because the download was in fact initiated")
+
+	r = gotStream.NewReader(ctx)
+	gotStream.Seal()
+
+	gotN, err = ioz.DrainClose(r)
+	require.Equal(t, len(sentBody), gotN)
+	tu.RequireTake(t, gotStream.Filled())
+	tu.RequireTake(t, gotStream.Done())
+
+	streamErr := gotStream.Err()
+	require.Error(t, streamErr)
+	require.True(t, errors.Is(err, streamErr))
 	t.Logf("got stream err: %v", err)
-	require.Error(t, err)
+
 	require.True(t, errors.Is(err, io.ErrUnexpectedEOF))
-	require.Equal(t, len(sentBody), stream.Size())
+	require.Equal(t, len(sentBody), gotStream.Size())
 
 	// Verify that the server hasn't updated the cache,
 	// by checking that the file modification timestamps
@@ -284,6 +300,4 @@ func TestCachePreservedOnFailedRefresh(t *testing.T) {
 	require.True(t, ioz.FileInfoEqual(fiBody1, fiBody2))
 	require.True(t, ioz.FileInfoEqual(fiHeader1, fiHeader2))
 	require.True(t, ioz.FileInfoEqual(fiChecksums1, fiChecksums2))
-
-	h.Reset()
 }

libsq/source/location/location.go

@@ -1,6 +1,9 @@
 // Package location contains functionality related to source location.
 package location
 
+// NOTE: This package contains code from several eras. There's a bunch of
+// overlap and duplication. It should be consolidated.
+
 import (
 	"net/url"
 	"path"
@@ -328,9 +331,9 @@ type Type string
 
 const (
 	TypeStdin   = "stdin"
-	TypeLocalFile  = "local_file"
+	TypeFile    = "local_file"
 	TypeSQL     = "sql"
-	TypeRemoteFile = "remote_file"
+	TypeHTTP    = "http_file"
 	TypeUnknown = "unknown"
 )
 
@@ -345,14 +348,14 @@ func TypeOf(loc string) Type {
 		return TypeSQL
 	case strings.HasPrefix(loc, "http://"),
 		strings.HasPrefix(loc, "https://"):
-		return TypeRemoteFile
+		return TypeHTTP
 	default:
 	}
 
 	if _, err := filepath.Abs(loc); err != nil {
 		return TypeUnknown
 	}
-	return TypeLocalFile
+	return TypeFile
 }
 
 // isHTTP tests if s is a well-structured HTTP or HTTPS url, and

testh/tu/tu.go

@@ -216,6 +216,26 @@ var (
 	_ AssertCompareFunc = require.Greater
 )
 
+// RequireNoTake fails if a value is taken from c.
+func RequireNoTake[C any](tb testing.TB, c <-chan C, msgAndArgs ...any) {
+	tb.Helper()
+	select {
+	case <-c:
+		require.Fail(tb, "unexpected take from channel", msgAndArgs...)
+	default:
+	}
+}
+
+// RequireTake fails if a value is not taken from c.
+func RequireTake[C any](tb testing.TB, c <-chan C, msgAndArgs ...any) {
+	tb.Helper()
+	select {
+	case <-c:
+	default:
+		require.Fail(tb, "unexpected failure to take from channel", msgAndArgs...)
+	}
+}
+
 // DirCopy copies the contents of sourceDir to a temp dir.
 // If keep is false, temp dir will be cleaned up on test exit.
 func DirCopy(tb testing.TB, sourceDir string, keep bool) (tmpDir string) {
@@ -356,6 +376,47 @@ func MustStat(tb testing.TB, fp string) os.FileInfo {
 	return fi
 }
 
+// MustDrain drains r, failing t on error. If arg cloze is true,
+// r is closed if it's an io.Closer, even if the drain fails.
+// FIXME: delete this func.
+func MustDrain(tb testing.TB, r io.Reader, cloze bool) {
+	tb.Helper()
+	_, cpErr := io.Copy(io.Discard, r)
+	if !cloze {
+		require.NoError(tb, cpErr)
+		return
+	}
+
+	var closeErr error
+	if rc, ok := r.(io.Closer); ok {
+		closeErr = rc.Close()
+	}
+
+	require.NoError(tb, cpErr)
+	require.NoError(tb, closeErr)
+}
+
+// MustDrainN is like MustDrain, but also reports the number of bytes
+// drained. If arg cloze is true, r is closed if it's an io.Closer,
+// even if the drain fails.
+func MustDrainN(tb testing.TB, r io.Reader, cloze bool) int {
+	tb.Helper()
+	n, cpErr := io.Copy(io.Discard, r)
+	if !cloze {
+		require.NoError(tb, cpErr)
+		return int(n)
+	}
+
+	var closeErr error
+	if rc, ok := r.(io.Closer); ok {
+		closeErr = rc.Close()
+	}
+
+	require.NoError(tb, cpErr)
+	require.NoError(tb, closeErr)
+	return int(n)
+}
+
 // TempDir is the standard means for obtaining a temp dir for tests.
 // If arg clean is true, the temp dir is created via t.TempDir, and
 // thus is deleted on test cleanup.