diff --git a/vortex-io/src/object_store/write.rs b/vortex-io/src/object_store/write.rs
index 3cb73cec331..1d7f42d5df9 100644
--- a/vortex-io/src/object_store/write.rs
+++ b/vortex-io/src/object_store/write.rs
@@ -21,10 +21,16 @@ use crate::VortexWrite;
 /// Adapter type to write data through a [`ObjectStore`] instance.
 ///
 /// After writing, the caller must make sure to call `shutdown`, in order to ensure the data is actually persisted.
+///
+/// A write that fails aborts the upload before returning, because S3 and GCS keep the parts that
+/// already landed when a [`MultipartUpload`] is dropped without one.
 pub struct ObjectStoreWrite {
     upload: Box<dyn MultipartUpload>,
     buffer: BytesMut,
     put_result: Option<PutResult>,
+    /// Set once the upload has been aborted. Aborting an already-aborted upload is
+    /// implementation-defined behaviour, so no path may do it twice.
+    aborted: bool,
 }
 
 const CHUNK_SIZE: usize = 16 * 1024 * 1024;
@@ -37,12 +43,27 @@ impl ObjectStoreWrite {
             upload,
             buffer: BytesMut::with_capacity(CHUNK_SIZE),
             put_result: None,
+            aborted: false,
         })
     }
 
     pub fn put_result(&self) -> Option<&PutResult> {
         self.put_result.as_ref()
     }
+
+    /// Abort the upload and hand back `error`, which is the one the caller needs to see.
+    ///
+    /// A failing abort is only logged: it cannot be recovered from here, and returning it instead
+    /// would hide why the write failed.
+    async fn abort_with(&mut self, error: io::Error) -> io::Error {
+        if !self.aborted {
+            self.aborted = true;
+            if let Err(abort_error) = self.upload.abort().await {
+                tracing::warn!("failed to abort multipart upload: {abort_error}");
+            }
+        }
+        error
+    }
 }
 
 impl VortexWrite for ObjectStoreWrite {
@@ -61,7 +82,9 @@ impl VortexWrite for ObjectStoreWrite {
             }
         }
 
-        parts.try_collect::<Vec<_>>().await?;
+        if let Err(error) = parts.try_collect::<Vec<_>>().await {
+            return Err(self.abort_with(error.into()).await);
+        }
 
         Ok(buffer)
     }
@@ -76,31 +99,45 @@ impl VortexWrite for ObjectStoreWrite {
             parts.push(part_fut);
         }
 
-        parts.try_collect::<Vec<_>>().await?;
+        if let Err(error) = parts.try_collect::<Vec<_>>().await {
+            return Err(self.abort_with(error.into()).await);
+        }
 
         Ok(())
     }
 
     async fn shutdown(&mut self) -> io::Result<()> {
+        // `flush` aborts the upload itself when it fails.
         self.flush().await?;
 
         if !self.buffer.is_empty() {
             let payload = std::mem::take(&mut self.buffer).freeze();
-            self.upload
-                .put_part(PutPayload::from_bytes(payload))
-                .await?;
+            let part = self.upload.put_part(PutPayload::from_bytes(payload));
+            if let Err(error) = part.await {
+                return Err(self.abort_with(error.into()).await);
+            }
         }
 
-        self.put_result = Some(self.upload.complete().await?);
-        Ok(())
+        let completed = self.upload.complete().await;
+        match completed {
+            Ok(put_result) => {
+                self.put_result = Some(put_result);
+                Ok(())
+            }
+            Err(error) => Err(self.abort_with(error.into()).await),
+        }
     }
 }
 
 #[cfg(test)]
 mod tests {
     use std::sync::Arc;
+    use std::sync::atomic::AtomicUsize;
+    use std::sync::atomic::Ordering;
 
+    use async_trait::async_trait;
     use object_store::ObjectStore;
+    use object_store::UploadPart;
     use object_store::local::LocalFileSystem;
     use object_store::memory::InMemory;
     use object_store::path::Path;
@@ -109,6 +146,94 @@ mod tests {
 
     use super::*;
 
+    /// A [`MultipartUpload`] that refuses every part and every completion, counting how often it
+    /// is aborted. S3 and GCS keep already-uploaded parts when an upload is dropped without an
+    /// abort, and neither `InMemory` nor `LocalFileSystem` exposes staged parts, so the count is
+    /// the only observable for that cleanup.
+    #[derive(Debug)]
+    struct RefusingUpload {
+        aborts: Arc<AtomicUsize>,
+    }
+
+    #[async_trait]
+    impl MultipartUpload for RefusingUpload {
+        fn put_part(&mut self, _data: PutPayload) -> UploadPart {
+            Box::pin(std::future::ready(Err(object_store::Error::Generic {
+                store: "refusing",
+                source: "put_part refused".into(),
+            })))
+        }
+
+        async fn complete(&mut self) -> object_store::Result<PutResult> {
+            Err(object_store::Error::Generic {
+                store: "refusing",
+                source: "complete refused".into(),
+            })
+        }
+
+        async fn abort(&mut self) -> object_store::Result<()> {
+            self.aborts.fetch_add(1, Ordering::SeqCst);
+            Ok(())
+        }
+    }
+
+    fn refusing_writer(aborts: &Arc<AtomicUsize>) -> ObjectStoreWrite {
+        ObjectStoreWrite {
+            upload: Box::new(RefusingUpload {
+                aborts: Arc::clone(aborts),
+            }),
+            buffer: BytesMut::with_capacity(CHUNK_SIZE),
+            put_result: None,
+            aborted: false,
+        }
+    }
+
+    /// The trailing part `shutdown` uploads is the one a small file goes out as, so its failure
+    /// must still abort.
+    #[tokio::test]
+    async fn shutdown_aborts_when_the_trailing_part_fails() -> anyhow::Result<()> {
+        let aborts = Arc::new(AtomicUsize::new(0));
+        let mut writer = refusing_writer(&aborts);
+
+        writer.write_all(vec![0u8; 8]).await?;
+        let error = writer.shutdown().await.expect_err("the part is refused");
+
+        assert!(error.to_string().contains("put_part refused"));
+        assert_eq!(aborts.load(Ordering::SeqCst), 1);
+        assert!(writer.put_result().is_none());
+        Ok(())
+    }
+
+    /// Nothing has been staged when only `complete` fails, but the upload is still open, so it
+    /// has to be aborted too.
+    #[tokio::test]
+    async fn shutdown_aborts_when_complete_fails() {
+        let aborts = Arc::new(AtomicUsize::new(0));
+        let mut writer = refusing_writer(&aborts);
+
+        let error = writer.shutdown().await.expect_err("completion is refused");
+
+        assert!(error.to_string().contains("complete refused"));
+        assert_eq!(aborts.load(Ordering::SeqCst), 1);
+        assert!(writer.put_result().is_none());
+    }
+
+    /// `flush` aborts on its own failure, and a caller that goes on to `shutdown` must not abort a
+    /// second time — `object_store` leaves that implementation-defined.
+    #[tokio::test]
+    async fn a_second_failure_does_not_abort_twice() -> anyhow::Result<()> {
+        let aborts = Arc::new(AtomicUsize::new(0));
+        let mut writer = refusing_writer(&aborts);
+
+        writer.write_all(vec![0u8; CHUNK_SIZE + 1]).await?;
+        writer.flush().await.expect_err("the part is refused");
+        assert_eq!(aborts.load(Ordering::SeqCst), 1);
+
+        writer.shutdown().await.expect_err("the part is refused");
+        assert_eq!(aborts.load(Ordering::SeqCst), 1);
+        Ok(())
+    }
+
     // Note: Concurrent writes test removed because &mut self in write_all already ensures
     // exclusive access. Multiple writers would need to be created with separate buffers,
     // which is not the intended use case.