Recently I have been working a lot with Cosmos and ran in to an issue when looking at deleting documents.
I need to delete around ~40 million documents in my Cosmos Container, I've looked around quite a bit and found a few options of which i have tried. two of the fastest of which I've tried are using a stored procedure within cosmos to delete records and using a bulk executor.
Both of these options have given subpar results compared to what I am looking for. I believe this should be obtainable within a couple hours but at the moment I am getting performance of around 1 hour per million recordsT
the two methods I used can also be seen here:
Stack Overflow Post on Document Deletion
My documents are about 35 keys long where half are string values and the other half are float/integer values, if that matters, and there are around 100k records per partition.
Here is are the two examples that I am using to attempt the deletion:
This first one is using C# and the documentation that helped me with this is here:
GitHub Documentation azure-cosmosdb-bulkexecutor-dotnet-getting-started
using System;
using System.Collections.Generic;
using System.Configuration;
using System.Diagnostics;
using System.Linq;
using System.Text;
using System.Threading;
using System.Threading.Tasks;
using Microsoft.Azure.Documents;
using Microsoft.Azure.Documents.Client;
using Microsoft.Azure.CosmosDB.BulkExecutor;
using Microsoft.Azure.CosmosDB.BulkExecutor.BulkImport;
using Microsoft.Azure.CosmosDB.BulkExecutor.BulkDelete;
namespace BulkDeleteSample
{
class Program
{
private static readonly string EndpointUrl = "xxxx";
private static readonly string AuthorizationKey = "xxxx";
private static readonly string DatabaseName = "xxxx";
private static readonly string CollectionName = "xxxx";
static ConnectionPolicy connectionPolicy = new ConnectionPolicy
{
ConnectionMode = ConnectionMode.Direct,
ConnectionProtocol = Protocol.Tcp
};
static async Task Main(string[] args)
{
DocumentClient client = new DocumentClient(new Uri(EndpointUrl), AuthorizationKey, connectionPolicy);
DocumentCollection dataCollection = GetCollectionIfExists(client, DatabaseName, CollectionName);
// Set retry options high during initialization (default values).
client.ConnectionPolicy.RetryOptions.MaxRetryWaitTimeInSeconds = 30;
client.ConnectionPolicy.RetryOptions.MaxRetryAttemptsOnThrottledRequests = 9;
BulkExecutor bulkExecutor = new BulkExecutor(client, dataCollection);
await bulkExecutor.InitializeAsync();
// Set retries to 0 to pass complete control to bulk executor.
client.ConnectionPolicy.RetryOptions.MaxRetryWaitTimeInSeconds = 0;
client.ConnectionPolicy.RetryOptions.MaxRetryAttemptsOnThrottledRequests = 0;
List<Tuple<string, string>> pkIdTuplesToDelete = new List<Tuple<string, string>>();
for (int i = 0; i < 99999; i++)
{
pkIdTuplesToDelete.Add(new Tuple<string, string>("1", i.ToString()));
}
BulkDeleteResponse bulkDeleteResponse = await bulkExecutor.BulkDeleteAsync(pkIdTuplesToDelete);
}
static DocumentCollection GetCollectionIfExists(DocumentClient client, string databaseName, string collectionName)
{
return client.CreateDocumentCollectionQuery(UriFactory.CreateDatabaseUri(databaseName))
.Where(c => c.Id == collectionName).AsEnumerable().FirstOrDefault();
}
}
}
The second one is using a stored procedure I found which delete data from a given partition using a query, of which I am running via a python notebook.
Here is the stored procedure:
/**
* A Cosmos DB stored procedure that bulk deletes documents for a given query.
* Note: You may need to execute this sproc multiple times (depending whether the sproc is able to delete every document within the execution timeout limit).
*
* @function
* @param {string} query - A query that provides the documents to be deleted (e.g. "SELECT c._self FROM c WHERE c.founded_year = 2008"). Note: For best performance, reduce the # of properties returned per document in the query to only what's required (e.g. prefer SELECT c._self over SELECT * )
* @returns {Object.<number, boolean>} Returns an object with the two properties:
* deleted - contains a count of documents deleted
* continuation - a boolean whether you should execute the sproc again (true if there are more documents to delete; false otherwise).
*/
function bulkDeleteSproc(query) {
var collection = getContext().getCollection();
var collectionLink = collection.getSelfLink();
var response = getContext().getResponse();
var responseBody = {
deleted: 0,
continuation: true
};
// Validate input.
if (!query) throw new Error("The query is undefined or null.");
tryQueryAndDelete();
// Recursively runs the query w/ support for continuation tokens.
// Calls tryDelete(documents) as soon as the query returns documents.
function tryQueryAndDelete(continuation) {
var requestOptions = {continuation: continuation};
var isAccepted = collection.queryDocuments(collectionLink, query, requestOptions, function (err, retrievedDocs, responseOptions) {
if (err) throw err;
if (retrievedDocs.length > 0) {
// Begin deleting documents as soon as documents are returned form the query results.
// tryDelete() resumes querying after deleting; no need to page through continuation tokens.
// - this is to prioritize writes over reads given timeout constraints.
tryDelete(retrievedDocs);
} else if (responseOptions.continuation) {
// Else if the query came back empty, but with a continuation token; repeat the query w/ the token.
tryQueryAndDelete(responseOptions.continuation);
} else {
// Else if there are no more documents and no continuation token - we are finished deleting documents.
responseBody.continuation = false;
response.setBody(responseBody);
}
});
// If we hit execution bounds - return continuation: true.
if (!isAccepted) {
response.setBody(responseBody);
}
}
// Recursively deletes documents passed in as an array argument.
// Attempts to query for more on empty array.
function tryDelete(documents) {
if (documents.length > 0) {
// Delete the first document in the array.
var isAccepted = collection.deleteDocument(documents[0]._self, {}, function (err, responseOptions) {
if (err) throw err;
responseBody.deleted++;
documents.shift();
// Delete the next document in the array.
tryDelete(documents);
});
// If we hit execution bounds - return continuation: true.
if (!isAccepted) {
response.setBody(responseBody);
}
} else {
// If the document array is empty, query for more documents.
tryQueryAndDelete();
}
}
}
I'm not sure if I am doing anything wrong or it the performance just isn't there with cosmos but I'm finding it quite difficult to achieve what I'm looking for, any advice is greatly appreciated.
