SentinelOne Singularity Operations Center - Cloud Funnel and the Filters That Decide What Leaves
Updated: 7 minutes ago

Before starts
Check Out old Series > https://www.cyberengage.org/courses-1/mastering-sentinelone%3A-a-comprehensive-guide-to-deep-visibility%2C-threat-hunting%2C-and-advanced-querying%22
This feature was available in the previous console as well, but I do not think I ever covered it, so I am covering it here.
Cloud Funnel sits under Singularity Data Lake. It takes your EDR and XDR telemetry and writes a copy of it into storage that you own. Compressed JSON, structured and enriched the same way Deep Visibility data is, so if you can read DV you can read this.
The reason you would want it: retention in the platform is whatever you pay for, retention in your own bucket is whatever you decide.
Add to that joining SentinelOne events against your other logs in your own SIEM, and that is the whole case for the feature.
So here you can decide whether you want to send all of your data, or only a little of it. That decision is the real content of this article, and most of it comes down to two filters I will get to further down.
First, go and check whether it is already on
Do you actually know whether Cloud Funnel is switched on in your tenant?

I did not. I went and looked properly for the first time last week and it was enabled on multiple sites.
And all sites were streaming everything. No query filter. No field selection. Every field of every event, leaving the platform continuously, into buckets nobody had looked at.
Nothing was broken. Nobody had made a mistake. It was working exactly as configured. The problem was that nobody had revisited the configuration since the day it was set up, and the default configuration is "send all of it".
So before you read the rest of this, go and open Singularity Data Lake → Cloud Funnel and find out which scopes have a configuration. It takes two minutes.
Where it can send
Five destinations:

Amazon S3 — from Union Square
Google Cloud Storage — from version W
Microsoft Azure Blob Storage container — from version Y SP4
Cloudflare R2 bucket
Google SecOps
Each wants something slightly different in the box.
Azure takes a Blob SAS URL, not a bucket name, so you need the Blob URL and the SAS token together in one string. R2 wants the S3 API string you generate from the R2 bucket's own settings, not the friendly name. Google SecOps wants a region, for example us, and a log type, where SENTINELONE_CF is the standard one.
Licensing is Singularity Complete plus the Cloud Funnel add-on, and it does not work in self-hosted environments at all.
Do the volume arithmetic before you enable anything
SentinelOne publish typical per-agent-per-day averages. Read the bottom half of this table slowly:
Agent type | Events per day, per agent |
Windows 22.3 and above | 90,000 |
Windows 22.2 and below | 140,000 |
macOS | 140,000 (around 280MB uncompressed) |
Linux agents | 500,000 |
Linux servers | 4,000,000 |
Kubernetes agents | 2,000,000 |
One Linux server produces more events per day than forty-four Windows endpoints. So the number you work out from a workstation estate stops being the number the moment somebody adds a Kubernetes cluster.
Multiply your own fleet out, multiply by your retention period, then look at your storage pricing.
Do it now rather than when the invoice arrives.
There is a built-in estimator on the configuration page, and from S-24.1.0 it shows you a projected volume when your selected scope is a Site. One warning from my own testing: when I drove the estimator through the API it returned the same account-wide total no matter which site I asked about. The console page may well behave properly, but check the number against something you trust before you plan a per-site rollout on it.
Scope and inheritance, which has a trap in it
Three levels: Global, then Account, then Site. Site-level configuration arrived at S-24.2.1.
Inheritance flows downward. Accounts inherit Global, Sites inherit their Account, and a child configuration overrides its parent for that child only.
You break inheritance to give one scope its own settings, and the page tells you which state you are in.
Now the trap. If you enable Cloud Funnel on an Account but no Site underneath it is enabled, no data flows. The configuration is valid, the bucket validates, everything looks right, nothing arrives. You need at least one Site switched on.
One timing quirk in the same area: an inheriting scope can take fifteen minutes before data shows up. Break the inheritance and it goes immediately.
Turning it on
Prerequisites: Cloud Funnel enabled with the right permissions, and Singularity Data Lake enabled — you need that second one for query filtering and for testing your filter in Event Search.
Permissions are four of their own under Cloud Funnel in the role: View, Edit, Delete, Create. Admin has them all. C-Level and Viewer can only look.

Set your storage up outside SentinelOne first. For S3 the permissions SentinelOne needs are List and Write.
On AWS, do not just open the bucket up. Create an IAM role with an external ID that SentinelOne gives you and let SentinelOne assume that role. It is more work once and it means the vendor does not hold standing bucket permissions.
Then in the console: pick the scope, pick the provider, enter the bucket, click Validate. That validation is better than most, because it only ever gives you one of three answers:

Given bucket does not exist — the name is wrong, compare it character by character
DataSet cannot access given destination — the bucket is there, but List and Write were not granted
Bucket was validated successfully — done
If it passes, make sure Enable is selected for Telemetry Streaming, and save.

Then wait, because configuration changes take up to 35 minutes to appear in your bucket. Onboarding, filter changes, anything. Steady-state delivery is much faster — events land within five minutes of the platform receiving them — but changes to the configuration itself are slow.
The filters, which is the real point
Two independent controls, plus a third thing that arrived recently. Think of it as rows and columns.
Rows: the Query Filter
Give Cloud Funnel a Singularity Data Lake query and only matching events stream. Available from platform X SP3.
Two rules.
It must be S1QL 2.0. And no PowerQuery commands — no columns, no group by. This is a filter, not a transformation.
The workflow here is genuinely good. There is a Help With Query Language link, an Event Search button that runs your query in the Data Lake so you can see what it actually catches, and a Validate button. Use the Event Search button before you save. Guessing at a query that decides what you will have in six months is not a plan.

You can also work backwards: if you already have a saved Event Search that captures what you care about, save that search directly as a Cloud Funnel filter.
Columns: Fields to include
Separately, choose which fields get sent. In the console it is Fields to include → Select Fields, with an Available Fields list, an Included Fields list, Add and Remove between them, and an All button for no filtering. Ctrl and Shift work for multi-select.


In the API it is desiredFields, taking XDR schema field names like activity_name. Leave it empty and you get every field — that is the default, and that is what the three sites I found were doing.
SentinelOne describe this one plainly as a cost saving measure, and they are right.
The third thing: EDR Slim Schema
New at S-26.3.6. A relational schema built for export that reduces the per-event payload with no data loss — smaller output, same information.
The catch: enable it and the field selection becomes fixed. So it is slim schema or hand-picked fields, not both.
It is mandatory for Google SecOps, optional everywhere else.
Which order to do it in
My advice, and it is the opposite of what you might expect from someone who does incident response: trim columns before you trim rows.
Field filtering costs you detail inside an event. Query filtering costs you whole events, and you cannot get back an event you never captured. Six months later, a thinner event is an inconvenience; a missing one ends the investigation.
So: slim schema if your version has it, or pick fields deliberately if not.
Only add a query filter if the volume is still unacceptable, and when you do, filter out noise you can positively identify rather than filtering down to what you think you will need. Those are different queries, and the second one is a prediction.
What it looks like in the bucket
Default layout:
s3://<bucket>/s1/cloud_funnel/<YYYY>/<MM>/<DD>/account_id=<account_id>Date partitioned, then by account — Hive-style, so pointing Athena or BigQuery at it is straightforward.
From S-24.2.5, sites that do not inherit their Account configuration get their own folder:
s3://<bucket>/s1/cloud_funnel/<YYYY>/<MM>/<DD>/account_id=<id>/site_id=<site_id>Inheriting sites keep the account-level path, and the change only takes effect for new non-inheriting sites or after you change something on an existing one. So an estate that has been running a while probably has both layouts in the same bucket. Worth knowing before you write the query that reads it.
Timings and limits in one place
Events reach your bucket up to 5 minutes after the platform receives them
Configuration changes take up to 35 minutes to take effect
Inherited scopes can take 15 minutes; break inheritance and it is immediate
Disable the toggle or let the licence lapse and you get a 24-hour grace period
On how many configurations you can have, the documentation says two different things. The FAQ says 500 configurations across all Sites and Scopes in an Account. The inheritance page says 20 Sites per Account. Those cannot both be the whole story, so if you are planning a wide rollout, confirm it with support before you design around 500 and find out the limit is 20.
What I would do on Monday morning
1. Find out which scopes have a configuration. Not which are licensed — which are configured and streaming.
2. For each one streaming, look at the filters. Empty desiredFields and no query means you are sending everything. Decide whether that is a choice or an accident.
3. Find out who owns the buckets and who pays for them. In my case they were set up by someone else, years ago, and the cost had never been attributed to anything.
4. Trim fields before you trim events. Slim schema if you have S-26.3.6.
5. Check your Linux and Kubernetes footprint against those per-agent numbers before adding scopes.
So what do I think
I like this one. It does a single job properly. The output matches Deep Visibility so there is nothing new to learn, the bucket layout is partitioned the way query engines want, the validation gives three unambiguous answers instead of a stack trace, and the AssumeRole option means you do not have to hand a vendor standing bucket permissions. Rows by query, columns by field, plus slim schema is exactly the right set of controls.
The complaints are at the edges. The documented limits contradict each other. The Account-enabled-but-no-Sites trap will cost somebody a day. And the estimator did not respect scope when I drove it through the API, which makes capacity planning harder than it should be.
But the thing I will remember from writing this is not any of that.
It is going looking for a feature I assumed we were not using and finding it had been quietly shipping full telemetry to three cloud buckets for over a year, every filter wide open, because that is the default and nobody went back to look.
That is not a SentinelOne problem. That is what happens to configuration nobody owns.
So go and check yours.
I will meet you in the next article. Till then, keep learning.
- written from hands-on work with the console, By Dean
--------------------------------------------------Cyberengage--------------------------------------



Comments