Allow recovery of poisoned objects via PUT/DELETE/DeleteObjects while preserving object-lock; implement race-free GetObject/HeadObject ?partNumber=N with correct zero-length-part response

crates/myfsio-server/src/handlers/mod.rs

@@ -164,32 +164,37 @@ async fn ensure_object_lock_allows_write(
     key: &str,
     headers: Option<&HeaderMap>,
 ) -> Result<(), Response> {
-    match state.storage.head_object(bucket, key).await {
-        Ok(_) => {
-            let metadata = match state.storage.get_object_metadata(bucket, key).await {
-                Ok(metadata) => metadata,
-                Err(err) => return Err(storage_err_response(err)),
-            };
-            let bypass_governance = headers
-                .and_then(|headers| {
-                    headers
-                        .get("x-amz-bypass-governance-retention")
-                        .and_then(|value| value.to_str().ok())
-                })
-                .map(|value| value.eq_ignore_ascii_case("true"))
-                .unwrap_or(false);
-            if let Err(message) = object_lock::can_delete_object(&metadata, bypass_governance) {
-                return Err(s3_error_response(S3Error::new(
-                    S3ErrorCode::AccessDenied,
-                    message,
-                )));
-            }
-            Ok(())
-        }
-        Err(myfsio_storage::error::StorageError::ObjectNotFound { .. }) => Ok(()),
-        Err(myfsio_storage::error::StorageError::DeleteMarker { .. }) => Ok(()),
-        Err(err) => Err(storage_err_response(err)),
+    let head_res = state.storage.head_object(bucket, key).await;
+    let needs_lock_check = match &head_res {
+        Ok(_) => true,
+        Err(myfsio_storage::error::StorageError::ObjectCorrupted { .. }) => true,
+        Err(myfsio_storage::error::StorageError::ObjectNotFound { .. }) => return Ok(()),
+        Err(myfsio_storage::error::StorageError::DeleteMarker { .. }) => return Ok(()),
+        Err(_) => false,
+    };
+    if !needs_lock_check {
+        return Err(storage_err_response(head_res.err().unwrap()));
     }
+
+    let metadata = match state.storage.get_object_metadata(bucket, key).await {
+        Ok(metadata) => metadata,
+        Err(err) => return Err(storage_err_response(err)),
+    };
+    let bypass_governance = headers
+        .and_then(|headers| {
+            headers
+                .get("x-amz-bypass-governance-retention")
+                .and_then(|value| value.to_str().ok())
+        })
+        .map(|value| value.eq_ignore_ascii_case("true"))
+        .unwrap_or(false);
+    if let Err(message) = object_lock::can_delete_object(&metadata, bypass_governance) {
+        return Err(s3_error_response(S3Error::new(
+            S3ErrorCode::AccessDenied,
+            message,
+        )));
+    }
+    Ok(())
 }
 
 async fn ensure_object_version_lock_allows_delete(
@@ -1609,6 +1614,13 @@ pub async fn get_object(
         .and_then(|v| v.to_str().ok())
         .map(|s| s.to_string());
 
+    if range_header.is_some() && query.part_number.is_some() {
+        return s3_error_response(S3Error::new(
+            S3ErrorCode::InvalidRequest,
+            "Cannot specify both Range and partNumber on the same request",
+        ));
+    }
+
     if let Some(ref range_str) = range_header {
         return range_get_handler(&state, &bucket, &key, range_str, &query, &headers).await;
     }
@@ -1643,6 +1655,40 @@ pub async fn get_object(
         Err(e) => return storage_err_response(e),
     };
 
+    if let Some(part_number) = query.part_number {
+        match resolve_part_view(&snap_meta, part_number) {
+            Ok(view) if view.multipart => {
+                if view.length == 0 {
+                    if let Some(resp) = evaluate_get_preconditions(&headers, &snap_meta) {
+                        let _ = tokio::fs::remove_file(&snap_link).await;
+                        return resp;
+                    }
+                    let _ = tokio::fs::remove_file(&snap_link).await;
+                    let mut h =
+                        build_part_response_headers(&key, &snap_meta, &view, &query);
+                    apply_user_metadata(&mut h, &snap_meta.metadata);
+                    return (StatusCode::PARTIAL_CONTENT, h).into_response();
+                }
+                let range_str = format!("bytes={}-{}", view.start, view.start + view.length - 1);
+                return serve_range_from_snapshot(
+                    &state,
+                    snap_link,
+                    snap_meta,
+                    &range_str,
+                    &query,
+                    &headers,
+                    Some(view.parts_count),
+                )
+                .await;
+            }
+            Ok(_) => {}
+            Err(resp) => {
+                let _ = tokio::fs::remove_file(&snap_link).await;
+                return resp;
+            }
+        }
+    }
+
     // Evaluate preconditions against the served snapshot's metadata. A HEAD
     // taken earlier could disagree with the snapshot if a concurrent PUT
     // landed in between, causing us to serve a body that doesn't satisfy
@@ -1870,6 +1916,21 @@ pub async fn head_object(
             if let Some(resp) = evaluate_get_preconditions(&headers, &meta) {
                 return resp;
             }
+
+            let part_view = match query.part_number {
+                Some(n) => match resolve_part_view(&meta, n) {
+                    Ok(v) => Some(v),
+                    Err(resp) => return resp,
+                },
+                None => None,
+            };
+
+            if let Some(view) = part_view.as_ref().filter(|v| v.multipart) {
+                let mut headers = build_part_response_headers(&key, &meta, view, &query);
+                apply_user_metadata(&mut headers, &meta.metadata);
+                return (StatusCode::PARTIAL_CONTENT, headers).into_response();
+            }
+
             let mut headers = HeaderMap::new();
             headers.insert("content-length", meta.size.to_string().parse().unwrap());
             if let Some(ref etag) = meta.etag {
@@ -1905,6 +1966,134 @@ pub async fn head_object(
     }
 }
 
+struct PartView {
+    start: u64,
+    length: u64,
+    parts_count: u32,
+    multipart: bool,
+}
+
+fn build_part_response_headers(
+    key: &str,
+    meta: &myfsio_common::types::ObjectMeta,
+    view: &PartView,
+    query: &ObjectQuery,
+) -> HeaderMap {
+    let mut headers = HeaderMap::new();
+    headers.insert("content-length", view.length.to_string().parse().unwrap());
+    if view.length > 0 {
+        headers.insert(
+            "content-range",
+            format!(
+                "bytes {}-{}/{}",
+                view.start,
+                view.start + view.length - 1,
+                meta.size
+            )
+            .parse()
+            .unwrap(),
+        );
+    }
+    if let Some(ref etag) = meta.etag {
+        headers.insert("etag", format!("\"{}\"", etag).parse().unwrap());
+    }
+    insert_content_type(&mut headers, key, meta.content_type.as_deref());
+    headers.insert(
+        "last-modified",
+        meta.last_modified
+            .format("%a, %d %b %Y %H:%M:%S GMT")
+            .to_string()
+            .parse()
+            .unwrap(),
+    );
+    headers.insert("accept-ranges", "bytes".parse().unwrap());
+    apply_stored_response_headers(&mut headers, &meta.internal_metadata);
+    if let Some(ref requested_version) = query.version_id {
+        if let Ok(value) = requested_version.parse() {
+            headers.insert("x-amz-version-id", value);
+        }
+    } else if let Some(ref vid) = meta.version_id {
+        if let Ok(value) = vid.parse() {
+            headers.insert("x-amz-version-id", value);
+        }
+    }
+    headers.insert(
+        "x-amz-mp-parts-count",
+        view.parts_count.to_string().parse().unwrap(),
+    );
+    apply_response_overrides(&mut headers, query);
+    headers
+}
+
+fn resolve_part_view(
+    meta: &myfsio_common::types::ObjectMeta,
+    part_number: u32,
+) -> Result<PartView, Response> {
+    if part_number < 1 {
+        return Err(s3_error_response(S3Error::new(
+            S3ErrorCode::InvalidArgument,
+            "partNumber must be >= 1",
+        )));
+    }
+
+    let etag = meta.etag.as_deref().unwrap_or("");
+    let is_multipart = myfsio_storage::fs_backend::is_multipart_etag(etag);
+
+    if !is_multipart {
+        if part_number == 1 {
+            return Ok(PartView {
+                start: 0,
+                length: meta.size,
+                parts_count: 1,
+                multipart: false,
+            });
+        }
+        return Err(s3_error_response(S3Error::new(
+            S3ErrorCode::InvalidPart,
+            format!(
+                "partNumber {} is out of range for a non-multipart object",
+                part_number
+            ),
+        )));
+    }
+
+    let part_sizes = match meta
+        .internal_metadata
+        .get(myfsio_storage::fs_backend::META_KEY_PART_SIZES)
+        .and_then(|raw| myfsio_storage::fs_backend::parse_part_sizes(raw))
+    {
+        Some(sizes) => sizes,
+        None => {
+            return Err(s3_error_response(S3Error::new(
+                S3ErrorCode::InvalidRequest,
+                "Object is multipart but has no recorded part-size manifest; \
+                 partNumber addressing is unavailable",
+            )));
+        }
+    };
+
+    let idx = (part_number as usize).saturating_sub(1);
+    if idx >= part_sizes.len() {
+        return Err(s3_error_response(S3Error::new(
+            S3ErrorCode::InvalidPart,
+            format!(
+                "partNumber {} exceeds the {} parts in this object",
+                part_number,
+                part_sizes.len()
+            ),
+        )));
+    }
+
+    let start: u64 = part_sizes.iter().take(idx).sum();
+    let length = part_sizes[idx];
+    Ok(PartView {
+        start,
+        length,
+        parts_count: part_sizes.len() as u32,
+        multipart: true,
+    })
+}
+
 async fn initiate_multipart_handler(state: &AppState, bucket: &str, key: &str) -> Response {
     match state.storage.initiate_multipart(bucket, key, None).await {
         Ok(upload_id) => {
@@ -2486,34 +2675,37 @@ async fn delete_objects_handler(state: &AppState, bucket: &str, body: Body) -> R
             async move {
                 let key = obj.key.clone();
                 let requested_vid = obj.version_id.clone();
+                let to_err = |err: myfsio_storage::error::StorageError| -> (String, String) {
+                    let s3err = S3Error::from(err);
+                    (s3err.code.as_str().to_string(), s3err.message)
+                };
+                let run_can_delete =
+                    |metadata: &HashMap<String, String>| -> Result<(), (String, String)> {
+                        object_lock::can_delete_object(metadata, false)
+                            .map_err(|m| (S3ErrorCode::AccessDenied.as_str().to_string(), m))
+                    };
                 let lock_check: Result<(), (String, String)> = match obj.version_id.as_deref() {
-                    Some(version_id) if version_id != "null" => match state
-                        .storage
-                        .get_object_version_metadata(&bucket, &obj.key, version_id)
-                        .await
-                    {
-                        Ok(metadata) => object_lock::can_delete_object(&metadata, false)
-                            .map_err(|m| (S3ErrorCode::AccessDenied.as_str().to_string(), m)),
-                        Err(err) => {
-                            let s3err = S3Error::from(err);
-                            Err((s3err.code.as_str().to_string(), s3err.message))
+                    Some(version_id) if version_id != "null" => {
+                        match state
+                            .storage
+                            .get_object_version_metadata(&bucket, &obj.key, version_id)
+                            .await
+                        {
+                            Ok(metadata) => run_can_delete(&metadata),
+                            Err(err) => Err(to_err(err)),
                         }
-                    },
+                    }
                     _ => match state.storage.head_object(&bucket, &obj.key).await {
-                        Ok(_) => match state.storage.get_object_metadata(&bucket, &obj.key).await {
-                            Ok(metadata) => object_lock::can_delete_object(&metadata, false)
-                                .map_err(|m| (S3ErrorCode::AccessDenied.as_str().to_string(), m)),
-                            Err(err) => {
-                                let s3err = S3Error::from(err);
-                                Err((s3err.code.as_str().to_string(), s3err.message))
+                        Ok(_)
+                        | Err(myfsio_storage::error::StorageError::ObjectCorrupted { .. }) => {
+                            match state.storage.get_object_metadata(&bucket, &obj.key).await {
+                                Ok(metadata) => run_can_delete(&metadata),
+                                Err(err) => Err(to_err(err)),
                             }
-                        },
+                        }
                         Err(myfsio_storage::error::StorageError::ObjectNotFound { .. }) => Ok(()),
                         Err(myfsio_storage::error::StorageError::DeleteMarker { .. }) => Ok(()),
-                        Err(err) => {
-                            let s3err = S3Error::from(err);
-                            Err((s3err.code.as_str().to_string(), s3err.message))
-                        }
+                        Err(err) => Err(to_err(err)),
                     },
                 };
 
@@ -2578,6 +2770,18 @@ async fn range_get_handler(
     range_str: &str,
     query: &ObjectQuery,
     headers: &HeaderMap,
+) -> Response {
+    range_get_handler_inner(state, bucket, key, range_str, query, headers, None).await
+}
+
+async fn range_get_handler_inner(
+    state: &AppState,
+    bucket: &str,
+    key: &str,
+    range_str: &str,
+    query: &ObjectQuery,
+    headers: &HeaderMap,
+    parts_count: Option<u32>,
 ) -> Response {
     let version_id = query
         .version_id
@@ -2607,6 +2811,21 @@ async fn range_get_handler(
         Err(e) => return storage_err_response(e),
     };
 
+    serve_range_from_snapshot(state, snap_link, meta, range_str, query, headers, parts_count).await
+}
+
+async fn serve_range_from_snapshot(
+    state: &AppState,
+    snap_link: std::path::PathBuf,
+    meta: myfsio_common::types::ObjectMeta,
+    range_str: &str,
+    query: &ObjectQuery,
+    headers: &HeaderMap,
+    parts_count: Option<u32>,
+) -> Response {
+    let key = meta.key.as_str();
+    let tmp_dir = state.config.storage_root.join(".myfsio.sys").join("tmp");
+
     if let Some(resp) = evaluate_get_preconditions(headers, &meta) {
         let _ = tokio::fs::remove_file(&snap_link).await;
         return resp;
@@ -2666,6 +2885,7 @@ async fn range_get_handler(
                         query,
                         Some(enc_info.algorithm.as_str()),
                         /* already_trimmed */ true,
+                        parts_count,
                     )
                     .await;
                 }
@@ -2720,6 +2940,7 @@ async fn range_get_handler(
         query,
         enc_header,
         /* already_trimmed */ false,
+        parts_count,
     )
     .await
 }
@@ -2735,6 +2956,7 @@ async fn stream_partial_content(
     query: &ObjectQuery,
     enc_header: Option<&str>,
     already_trimmed: bool,
+    parts_count: Option<u32>,
 ) -> Response {
     let length = end - start + 1;
 
@@ -2789,6 +3011,10 @@ async fn stream_partial_content(
 
     apply_response_overrides(&mut headers, query);
 
+    if let Some(count) = parts_count {
+        headers.insert("x-amz-mp-parts-count", count.to_string().parse().unwrap());
+    }
+
     (StatusCode::PARTIAL_CONTENT, headers, body).into_response()
 }
 

crates/myfsio-storage/src/fs_backend.rs

@@ -21,6 +21,34 @@ pub const META_KEY_CORRUPTED: &str = "__corrupted__";
 pub const META_KEY_CORRUPTED_AT: &str = "__corrupted_at__";
 pub const META_KEY_CORRUPTION_DETAIL: &str = "__corruption_detail__";
 pub const META_KEY_QUARANTINE_PATH: &str = "__quarantine_path__";
+pub const META_KEY_PART_SIZES: &str = "__part_sizes__";
+
+pub fn encode_part_sizes(sizes: &[u64]) -> String {
+    let mut out = String::with_capacity(sizes.len() * 8);
+    for (i, s) in sizes.iter().enumerate() {
+        if i > 0 {
+            out.push(',');
+        }
+        out.push_str(&s.to_string());
+    }
+    out
+}
+
+pub fn parse_part_sizes(raw: &str) -> Option<Vec<u64>> {
+    let mut out = Vec::new();
+    for tok in raw.split(',') {
+        let tok = tok.trim();
+        if tok.is_empty() {
+            return None;
+        }
+        out.push(tok.parse::<u64>().ok()?);
+    }
+    if out.is_empty() {
+        None
+    } else {
+        Some(out)
+    }
+}
 
 pub fn metadata_is_corrupted(meta: &HashMap<String, String>) -> bool {
     meta.get(META_KEY_CORRUPTED)
@@ -2829,6 +2857,12 @@ impl crate::traits::StorageEngine for FsStorageBackend {
                     self.delete_metadata_sync(bucket, key)
                         .map_err(StorageError::Io)?;
                     Self::cleanup_empty_parents(&path, &bucket_path);
+                } else {
+                    let stored_meta = self.read_metadata_sync(bucket, key);
+                    if !stored_meta.is_empty() {
+                        self.delete_metadata_sync(bucket, key)
+                            .map_err(StorageError::Io)?;
+                    }
                 }
                 let dm_version_id = self
                     .write_delete_marker_sync(bucket, key)
@@ -2842,6 +2876,17 @@ impl crate::traits::StorageEngine for FsStorageBackend {
             }
 
             if !path.exists() {
+                let stored_meta = self.read_metadata_sync(bucket, key);
+                if !stored_meta.is_empty() {
+                    self.delete_metadata_sync(bucket, key)
+                        .map_err(StorageError::Io)?;
+                    self.invalidate_bucket_caches(bucket);
+                    return Ok(DeleteOutcome {
+                        version_id: None,
+                        is_delete_marker: false,
+                        existed: true,
+                    });
+                }
                 return Ok(DeleteOutcome::default());
             }
 
@@ -3353,47 +3398,52 @@ impl crate::traits::StorageEngine for FsStorageBackend {
 
         // Assemble parts on a blocking thread using std::fs, large buffers,
         // and a single writer flush — no per-chunk runtime crossings.
-        let assemble_res = tokio::task::spawn_blocking(move || -> StorageResult<(String, u64)> {
-            use std::io::{BufReader, BufWriter, Read, Write};
-            let out_raw = std::fs::File::create(&tmp_path_owned).map_err(StorageError::Io)?;
-            let mut out_file = BufWriter::with_capacity(chunk_size * 4, out_raw);
-            let mut md5_digest_concat = Vec::with_capacity(part_infos.len() * 16);
-            let mut total_size: u64 = 0;
-            let mut buf = vec![0u8; chunk_size];
+        let assemble_res =
+            tokio::task::spawn_blocking(move || -> StorageResult<(String, u64, Vec<u64>)> {
+                use std::io::{BufReader, BufWriter, Read, Write};
+                let out_raw = std::fs::File::create(&tmp_path_owned).map_err(StorageError::Io)?;
+                let mut out_file = BufWriter::with_capacity(chunk_size * 4, out_raw);
+                let mut md5_digest_concat = Vec::with_capacity(part_infos.len() * 16);
+                let mut total_size: u64 = 0;
+                let mut part_sizes: Vec<u64> = Vec::with_capacity(part_infos.len());
+                let mut buf = vec![0u8; chunk_size];
 
-            for part_info in &part_infos {
-                let part_file =
-                    upload_dir_owned.join(format!("part-{:05}.part", part_info.part_number));
-                if !part_file.exists() {
-                    return Err(StorageError::InvalidObjectKey(format!(
-                        "Part {} not found",
-                        part_info.part_number
-                    )));
-                }
-                let reader = std::fs::File::open(&part_file).map_err(StorageError::Io)?;
-                let mut reader = BufReader::with_capacity(chunk_size, reader);
-                let mut part_hasher = Md5::new();
-                loop {
-                    let n = reader.read(&mut buf).map_err(StorageError::Io)?;
-                    if n == 0 {
-                        break;
+                for part_info in &part_infos {
+                    let part_file =
+                        upload_dir_owned.join(format!("part-{:05}.part", part_info.part_number));
+                    if !part_file.exists() {
+                        return Err(StorageError::InvalidObjectKey(format!(
+                            "Part {} not found",
+                            part_info.part_number
+                        )));
                     }
-                    part_hasher.update(&buf[..n]);
-                    out_file.write_all(&buf[..n]).map_err(StorageError::Io)?;
-                    total_size += n as u64;
+                    let reader = std::fs::File::open(&part_file).map_err(StorageError::Io)?;
+                    let mut reader = BufReader::with_capacity(chunk_size, reader);
+                    let mut part_hasher = Md5::new();
+                    let mut part_bytes: u64 = 0;
+                    loop {
+                        let n = reader.read(&mut buf).map_err(StorageError::Io)?;
+                        if n == 0 {
+                            break;
+                        }
+                        part_hasher.update(&buf[..n]);
+                        out_file.write_all(&buf[..n]).map_err(StorageError::Io)?;
+                        total_size += n as u64;
+                        part_bytes += n as u64;
+                    }
+                    md5_digest_concat.extend_from_slice(&part_hasher.finalize());
+                    part_sizes.push(part_bytes);
                 }
-                md5_digest_concat.extend_from_slice(&part_hasher.finalize());
-            }
 
-            out_file.flush().map_err(StorageError::Io)?;
-            let mut composite_hasher = Md5::new();
-            composite_hasher.update(&md5_digest_concat);
-            let etag = format!("{:x}-{}", composite_hasher.finalize(), part_infos.len());
-            Ok((etag, total_size))
-        })
-        .await;
+                out_file.flush().map_err(StorageError::Io)?;
+                let mut composite_hasher = Md5::new();
+                composite_hasher.update(&md5_digest_concat);
+                let etag = format!("{:x}-{}", composite_hasher.finalize(), part_infos.len());
+                Ok((etag, total_size, part_sizes))
+            })
+            .await;
 
-        let (etag, total_size) = match assemble_res {
+        let (etag, total_size, part_sizes) = match assemble_res {
             Ok(Ok(v)) => v,
             Ok(Err(e)) => {
                 let _ = std::fs::remove_file(&tmp_path);
@@ -3408,6 +3458,9 @@ impl crate::traits::StorageEngine for FsStorageBackend {
             }
         };
 
+        let mut metadata = metadata;
+        metadata.insert(META_KEY_PART_SIZES.to_string(), encode_part_sizes(&part_sizes));
+
         // Commit to the destination key atomically under its write lock.
         // Lock acquisition happens inside run_blocking so the wait runs under
         // block_in_place rather than parking the async worker.
@@ -4033,6 +4086,117 @@ mod tests {
         );
     }
 
+    #[test]
+    fn test_part_sizes_roundtrip() {
+        let sizes = vec![5_242_880, 5_242_880, 5_242_880, 12_345];
+        let encoded = encode_part_sizes(&sizes);
+        assert_eq!(encoded, "5242880,5242880,5242880,12345");
+        let parsed = parse_part_sizes(&encoded).unwrap();
+        assert_eq!(parsed, sizes);
+        assert!(parse_part_sizes("").is_none());
+        assert!(parse_part_sizes(",,,").is_none());
+        assert!(parse_part_sizes("abc").is_none());
+        assert!(parse_part_sizes("123,abc").is_none());
+        assert!(parse_part_sizes(" ").is_none());
+    }
+
+    #[tokio::test]
+    async fn test_delete_object_clears_poisoned_metadata() {
+        let (_dir, backend) = create_test_backend();
+        backend.create_bucket("test-bucket").await.unwrap();
+
+        let data: AsyncReadStream = Box::pin(std::io::Cursor::new(b"will rot".to_vec()));
+        backend
+            .put_object("test-bucket", "rot.txt", data, None)
+            .await
+            .unwrap();
+
+        let mut meta = backend
+            .get_object_metadata("test-bucket", "rot.txt")
+            .await
+            .unwrap();
+        meta.insert(META_KEY_CORRUPTED.to_string(), "true".to_string());
+        backend
+            .put_object_metadata("test-bucket", "rot.txt", &meta)
+            .await
+            .unwrap();
+
+        let live_path = backend
+            .get_object_path("test-bucket", "rot.txt")
+            .await
+            .unwrap();
+        std::fs::remove_file(&live_path).unwrap();
+
+        backend
+            .delete_object("test-bucket", "rot.txt")
+            .await
+            .unwrap();
+
+        match backend.head_object("test-bucket", "rot.txt").await {
+            Err(StorageError::ObjectNotFound { .. }) => {}
+            other => panic!(
+                "after DELETE on a poisoned/quarantined object, HEAD should be ObjectNotFound, got {:?}",
+                other
+            ),
+        }
+
+        let leftover = backend
+            .get_object_metadata("test-bucket", "rot.txt")
+            .await
+            .unwrap();
+        assert!(
+            leftover.is_empty(),
+            "metadata sidecar must be cleared after DELETE on poisoned object"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_complete_multipart_persists_part_sizes() {
+        let (_dir, backend) = create_test_backend();
+        backend.create_bucket("mp-bucket").await.unwrap();
+
+        let upload_id = backend
+            .initiate_multipart("mp-bucket", "obj.bin", None)
+            .await
+            .unwrap();
+
+        let part1: AsyncReadStream = Box::pin(std::io::Cursor::new(vec![b'A'; 1024]));
+        backend
+            .upload_part("mp-bucket", &upload_id, 1, part1)
+            .await
+            .unwrap();
+        let part2: AsyncReadStream = Box::pin(std::io::Cursor::new(vec![b'B'; 512]));
+        backend
+            .upload_part("mp-bucket", &upload_id, 2, part2)
+            .await
+            .unwrap();
+
+        let parts = vec![
+            PartInfo {
+                part_number: 1,
+                etag: String::new(),
+            },
+            PartInfo {
+                part_number: 2,
+                etag: String::new(),
+            },
+        ];
+        let obj = backend
+            .complete_multipart("mp-bucket", &upload_id, &parts)
+            .await
+            .unwrap();
+        assert_eq!(obj.size, 1536);
+
+        let stored = backend
+            .get_object_metadata("mp-bucket", "obj.bin")
+            .await
+            .unwrap();
+        let raw = stored
+            .get(META_KEY_PART_SIZES)
+            .expect("part sizes must be persisted on completion");
+        assert_eq!(parse_part_sizes(raw).unwrap(), vec![1024u64, 512u64]);
+    }
+
     #[tokio::test]
     async fn test_put_clears_poison_flag() {
         let (_dir, backend) = create_test_backend();