diff --git a/gcsfs/core.py b/gcsfs/core.py
index f425afc0..f6f50036 100644
--- a/gcsfs/core.py
+++ b/gcsfs/core.py
@@ -25,7 +25,7 @@
 from .checkers import get_consistency_checker
 from .credentials import GoogleCredentials
 from .inventory_report import InventoryReport
-from .retry import retry_request, validate_response
+from .retry import errs, retry_request, validate_response
 
 logger = logging.getLogger("gcsfs")
 
@@ -1168,6 +1168,8 @@ async def _rm_file(self, path, **kwargs):
             await self._rmdir(path)
 
     async def _rm_files(self, paths):
+        import random
+
         template = (
             "\n--===============7330845974216740156==\n"
             "Content-Type: application/http\n"
@@ -1178,46 +1180,64 @@ async def _rm_files(self, paths):
             "accept: application/json\ncontent-length: 0\n"
         )
         errors = []
+        success = []
         # Splitting requests into 100 chunk batches
         # See https://cloud.google.com/storage/docs/batch
-        for chunk in _chunks(paths, 100):
-            parts = []
-            for i, p in enumerate(chunk):
-                bucket, key, generation = self.split_path(p)
-                query = f"?generation={generation}" if generation else ""
-                parts.append(
-                    template.format(
-                        i=i + 1,
-                        bucket=quote(bucket),
-                        key=quote(key),
-                        query=query,
+        for retry in range(1, 6):
+            for chunk in _chunks(paths, 100):
+                parts = []
+                for i, p in enumerate(chunk):
+                    bucket, key, generation = self.split_path(p)
+                    query = f"?generation={generation}" if generation else ""
+                    parts.append(
+                        template.format(
+                            i=i + 1,
+                            bucket=quote(bucket),
+                            key=quote(key),
+                            query=query,
+                        )
                     )
+                body = "".join(parts)
+                headers, content = await self._call(
+                    "POST",
+                    f"{self._location}/batch/storage/v1",
+                    headers={
+                        "Content-Type": 'multipart/mixed; boundary="=========='
+                        '=====7330845974216740156=="'
+                    },
+                    data=body + "\n--===============7330845974216740156==--",
                 )
-            body = "".join(parts)
-            headers, content = await self._call(
-                "POST",
-                f"{self._location}/batch/storage/v1",
-                headers={
-                    "Content-Type": 'multipart/mixed; boundary="=========='
-                    '=====7330845974216740156=="'
-                },
-                data=body + "\n--===============7330845974216740156==--",
-            )
 
-            boundary = headers["Content-Type"].split("=", 1)[1]
-            parents = [self._parent(p) for p in paths]
-            [self.invalidate_cache(parent) for parent in parents + list(paths)]
-            txt = content.decode()
-            if any(
-                not ("200 OK" in c or "204 No Content" in c)
-                for c in txt.split(boundary)[1:-1]
-            ):
-                pattern = '"message": "([^"]+)"'
-                out = set(re.findall(pattern, txt))
-                errors.extend(out)
+                boundary = headers["Content-Type"].split("=", 1)[1]
+                parents = [self._parent(p) for p in paths]
+                [self.invalidate_cache(parent) for parent in parents + list(paths)]
+                txt = content.decode()
+                responses = txt.split(boundary)[1:-1]
+                remaining = []
+                for path, response in zip(paths, responses):
+                    m = re.search("HTTP/[0-9.]+ ([0-9]+)", response)
+                    code = int(m.groups()[0]) if m else None
+                    if code in [200, 204]:
+                        success.append(path)
+                    elif code in errs:
+                        remaining.append(path)
+                    else:
+                        msg = re.search("({.*})", response.replace("\n", ""))
+                        if msg:
+                            errors.append(json.loads(msg.groups()[0])["error"])
+                        else:
+                            errors.append((path, code))
+                if errors:
+                    break
+                if remaining:
+                    paths = remaining
+                    await asyncio.sleep(min(random.random() + 2 ** (retry - 1), 32))
+                else:
+                    break
 
         if errors:
             raise OSError(errors)
+        return success
 
     @property
     def on_google(self):
diff --git a/gcsfs/retry.py b/gcsfs/retry.py
index 4bb860ba..793eaba5 100644
--- a/gcsfs/retry.py
+++ b/gcsfs/retry.py
@@ -58,14 +58,17 @@ class ChecksumError(Exception):
 )
 
 
+errs = list(range(500, 505)) + [
+    # Request Timeout
+    408,
+    # Too Many Requests
+    429,
+]
+
+
 def is_retriable(exception):
     """Returns True if this exception is retriable."""
-    errs = list(range(500, 505)) + [
-        # Request Timeout
-        408,
-        # Too Many Requests
-        429,
-    ]
+
     errs += [str(e) for e in errs]
     if isinstance(exception, HttpError):
         return exception.code in errs