aryn-ai · dhruvkaliraman7 · Feb 6, 2025 · Jan 26, 2025 · Jan 27, 2025 · Feb 5, 2025
diff --git a/lib/sycamore/sycamore/connectors/base_writer.py b/lib/sycamore/sycamore/connectors/base_writer.py
@@ -33,6 +33,15 @@ def create_target_idempotent(self, target_params: "BaseDBWriter.TargetParams"):
         def get_existing_target_params(self, target_params: "BaseDBWriter.TargetParams") -> "BaseDBWriter.TargetParams":
             pass
 
+        def reliability_assertor(self, target_params: "BaseDBWriter.TargetParams"):
+            """
+            Method to verify that all documents were successfully written when reliability mode is enabled.
+
+            Raises:
+                NotImplementedError: If the implementing class doesn't support reliability checks
+            """
+            raise NotImplementedError("This writer does not support reliability checks")
+
         def close(self):
             pass
 

diff --git a/lib/sycamore/sycamore/connectors/opensearch/opensearch_writer.py b/lib/sycamore/sycamore/connectors/opensearch/opensearch_writer.py
@@ -16,6 +16,9 @@
     DEFAULT_RECORD_PROPERTIES,
 )
 from sycamore.utils.import_utils import requires_modules
+from sycamore.plan_nodes import Node
+from sycamore.docset import DocSet
+from sycamore.context import Context
 
 if typing.TYPE_CHECKING:
     from opensearchpy import OpenSearch
@@ -42,6 +45,7 @@ class OpenSearchWriterClientParams(BaseDBWriter.ClientParams):
 @dataclass
 class OpenSearchWriterTargetParams(BaseDBWriter.TargetParams):
     index_name: str
+    _doc_count: int = 0
     settings: dict[str, Any] = field(default_factory=lambda: {"index.knn": True})
     mappings: dict[str, Any] = field(
         default_factory=lambda: {
@@ -92,6 +96,61 @@ def compatible_with(self, other: "BaseDBWriter.TargetParams") -> bool:
         other_flat_mappings = dict(flatten_data(other.mappings))
         return check_dictionary_compatibility(my_flat_mappings, other_flat_mappings)
 
+    @classmethod
+    def from_write_args(
+        cls,
+        index_name: str,
+        plan: Node,
+        context: Context,
+        reliability_rewriter: bool,
+        execute: bool,
+        insert_settings: Optional[dict] = None,
+        index_settings: Optional[dict] = None,
+    ) -> "OpenSearchWriterTargetParams":
+        """
+        Build OpenSearchWriterTargetParams from write operation arguments.
+
+        Args:
+            index_name: Name of the OpenSearch index
+            plan: The execution plan Node
+            context: The execution Context
+            reliability_rewriter: Whether to enable reliability rewriter mode
+            execute: Whether to execute the pipeline immediately
+            insert_settings: Optional settings for data insertion
+            index_settings: Optional index configuration settings
+
+        Returns:
+            OpenSearchWriterTargetParams configured with the provided settings
+
+        Raises:
+            AssertionError: If reliability_rewriter conditions are not met
+        """
+        target_params_dict: dict[str, Any] = {
+            "index_name": index_name,
+            "_doc_count": 0,
+        }
+
+        if reliability_rewriter:
+            from sycamore.materialize import Materialize
+
+            assert execute, "Reliability rewriter requires execute to be True"
+            assert (
+                type(plan) == Materialize
+            ), "The first node must be a materialize node for reliability rewriter to work"
+            assert not plan.children[
+                0
+            ], "Pipeline should only have read materialize and write nodes for reliability rewriter to work"
+            target_params_dict["_doc_count"] = DocSet(context, plan).count()
+
+        if insert_settings:
+            target_params_dict["insert_settings"] = insert_settings
+
+        if index_settings:
+            target_params_dict["settings"] = index_settings.get("body", {}).get("settings", {})
+            target_params_dict["mappings"] = index_settings.get("body", {}).get("mappings", {})
+
+        return cls(**target_params_dict)
+
 
 class OpenSearchWriterClient(BaseDBWriter.Client):
     def __init__(self, os_client: "OpenSearch"):
@@ -187,6 +246,8 @@ def _string_values_to_python_types(obj: Any):
                     return obj
             return obj
 
+        # TODO: Convert OpenSearchWriterTargetParams to pydantic model
+
         assert isinstance(
             target_params, OpenSearchWriterTargetParams
         ), f"Provided target_params was not of type OpenSearchWriterTargetParams:\n{target_params}"
@@ -196,7 +257,27 @@ def _string_values_to_python_types(obj: Any):
         assert isinstance(mappings, dict)
         settings = _string_values_to_python_types(response.get(index_name, {}).get("settings", {}))
         assert isinstance(settings, dict)
-        return OpenSearchWriterTargetParams(index_name=index_name, mappings=mappings, settings=settings)
+        _doc_count = target_params._doc_count
+        assert isinstance(_doc_count, int)
+        return OpenSearchWriterTargetParams(
+            index_name=index_name,
+            mappings=mappings,
+            settings=settings,
+            _doc_count=_doc_count,
+        )
+
+    def reliability_assertor(self, target_params: BaseDBWriter.TargetParams):
+        assert isinstance(
+            target_params, OpenSearchWriterTargetParams
+        ), f"Provided target_params was not of type OpenSearchWriterTargetParams:\n{target_params}"
+        log.info("Flushing index...")
+        self._client.indices.flush(index=target_params.index_name, params={"timeout": 300})
+        log.info("Done flushing index.")
+        indices = self._client.cat.indices(index=target_params.index_name, format="json")
+        assert len(indices) == 1, f"Expected 1 index, found {len(indices)}"
+        num_docs = int(indices[0]["docs.count"])
+        log.info(f"{num_docs} chunks written in index {target_params.index_name}")
+        assert num_docs == target_params._doc_count, f"Expected {target_params._doc_count} docs, found {num_docs}"
 
 
 @dataclass

diff --git a/lib/sycamore/sycamore/tests/integration/connectors/opensearch/test_opensearch_read.py b/lib/sycamore/sycamore/tests/integration/connectors/opensearch/test_opensearch_read.py
@@ -248,6 +248,56 @@ def remove_reconstruct_doc_property(doc: Document):
         # Clean up
         os_client.indices.delete(setup_index, ignore_unavailable=True)
 
+    def test_write_with_reliability(self, setup_index, os_client, exec_mode):
+        """
+        Validates that when materialized pickle outputs are deleted, the index is rewritten
+        with the correct (reduced) number of chunks.
+        """
+        with tempfile.TemporaryDirectory() as tmpdir1:
+            path = str(TEST_DIR / "resources/data/pdfs/Ray.pdf")
+            context = sycamore.init(exec_mode=exec_mode)
+
+            # 2 docs for ray execution
+            (
+                context.read.binary([path, path], binary_format="pdf")
+                .partition(ArynPartitioner(aryn_api_key=ARYN_API_KEY))
+                .explode()
+                .materialize(path=tmpdir1)
+                .execute()
+            )
+
+            (
+                context.read.materialize(tmpdir1).write.opensearch(
+                    os_client_args=TestOpenSearchRead.OS_CLIENT_ARGS,
+                    index_name=setup_index,
+                    index_settings=TestOpenSearchRead.INDEX_SETTINGS,
+                    reliability_rewriter=True,
+                )
+            )
+            os_client.indices.refresh(setup_index)
+            count = get_doc_count(os_client, setup_index)
+
+            # Delete 1 pickle file to make sure reliability rewriter works
+            pickle_files = [f for f in os.listdir(tmpdir1) if f.endswith(".pickle")]
+            assert pickle_files, "No pickle files found in materialized directory"
+            os.remove(os.path.join(tmpdir1, pickle_files[0]))
+
+            # Delete and recreate the index - should have fewer chunks
+            (
+                context.read.materialize(tmpdir1).write.opensearch(
+                    os_client_args=TestOpenSearchRead.OS_CLIENT_ARGS,
+                    index_name=setup_index,
+                    index_settings=TestOpenSearchRead.INDEX_SETTINGS,
+                    reliability_rewriter=True,
+                )
+            )
+            os_client.indices.refresh(setup_index)
+            re_count = get_doc_count(os_client, setup_index)
+
+        # Verify document count is reduced
+        assert count - 1 == re_count, f"Expected {count} documents, found {re_count}"
+        os_client.indices.delete(setup_index)
+
     def _test_ingest_and_read_via_docid_reconstructor(self, setup_index, os_client, cache_dir):
         """
         Validates data is readable from OpenSearch, and that we can rebuild processed Sycamore documents.

diff --git a/lib/sycamore/sycamore/writer.py b/lib/sycamore/sycamore/writer.py
@@ -22,6 +22,7 @@
     from neo4j import Auth
     from neo4j.auth_management import AuthManager
 
+from sycamore.connectors.base_writer import BaseDBWriter
 
 logger = logging.getLogger(__name__)
 
@@ -47,6 +48,7 @@ def opensearch(
         index_settings: dict,
         insert_settings: Optional[dict] = None,
         execute: bool = True,
+        reliability_rewriter: bool = False,
         **kwargs,
     ) -> Optional["DocSet"]:
         """Writes the content of the DocSet into the specified OpenSearch index.
@@ -103,9 +105,9 @@ def opensearch(
         """
 
         from sycamore.connectors.opensearch import (
-            OpenSearchWriter,
             OpenSearchWriterClientParams,
             OpenSearchWriterTargetParams,
+            OpenSearchWriter,
         )
         from typing import Any
         import copy
@@ -137,23 +139,30 @@ def _convert_to_host_port_list(hostlist: Any) -> list[HostAndPort]:
             os_client_args["hosts"] = _convert_to_host_port_list(hosts)
         client_params = OpenSearchWriterClientParams(**os_client_args)
 
-        target_params: OpenSearchWriterTargetParams
-        target_params_dict: dict[str, Any] = {"index_name": index_name}
-        if insert_settings:
-            target_params_dict["insert_settings"] = insert_settings
-        if index_settings:
-            target_params_dict["settings"] = index_settings.get("body", {}).get("settings", {})
-            target_params_dict["mappings"] = index_settings.get("body", {}).get("mappings", {})
-        target_params = OpenSearchWriterTargetParams(**target_params_dict)
+        target_params = OpenSearchWriterTargetParams.from_write_args(
+            index_name=index_name,
+            plan=self.plan,
+            context=self.context,
+            reliability_rewriter=reliability_rewriter,
+            execute=execute,
+            insert_settings=insert_settings,
+            index_settings=index_settings,
+        )
         os = OpenSearchWriter(
             self.plan, client_params=client_params, target_params=target_params, name="OsrchWrite", **kwargs
         )
+        client = None
+        if reliability_rewriter:
+            client = os.Client.from_client_params(client_params)
+            if client._client.indices.exists(index=index_name):
+                logger.info(f"\n\nWARNING WARNING WARNING: Deleting existing index {index_name}\n\n")
+                client._client.indices.delete(index=index_name)
 
         # We will probably want to break this at some point so that write
         # doesn't execute automatically, and instead you need to say something
         # like docset.write.opensearch().execute(), allowing sensible writes
         # to multiple locations and post-write operations.
-        return self._maybe_execute(os, execute)
+        return self._maybe_execute(os, execute, client)
 
     @requires_modules(["weaviate", "weaviate.collections.classes.config"], extra="weaviate")
     def weaviate(
@@ -849,10 +858,16 @@ def aryn(
 
         return self._maybe_execute(ds, True)
 
-    def _maybe_execute(self, node: Node, execute: bool) -> Optional[DocSet]:
+    def _maybe_execute(
+        self, node: Node, execute: bool, client: Optional[BaseDBWriter.Client] = None
+    ) -> Optional[DocSet]:
         ds = DocSet(self.context, node)
         if not execute:
             return ds
 
         ds.execute()
+
+        if client is not None:
+            if type(node) == BaseDBWriter:
+                client.reliability_assertor(node._target_params)
         return None