Re: Encryption of Data Specific to a Tenant in PostgreSQL database | General Idea

[Rob Sargent <robjsargent@xxxxxxxxx>]

On Feb 10, 2021, at 6:45 PM, Tim Cross <theophilusx@xxxxxxxxx> wrote:

Jagmohan Kaintura <jagmohan@xxxxxxxxxxxxxx> writes:

HI All,

For POstgreSQL database to store data for multiple tenants, the approach
decided was to have
Shared Database (Holding data for all tenants)
     => Data would be segregated on basis of some additional column
(tennatid,different tenants having different tenantId)
          => Data would be accessed through Views on the basis of tenantId
value.

This is the basic process of most of the customers who are trying to
implement multiple tenants in PostgreSQL, rather than choosing
separate databases for each tenant.

Now we need to encrypt the data related to a tenantId, so that now one
knows this data belongs to which tenant even from Operations group.
Is there a method in POstgreSQL for encrypting data with different keys
with respect to different values in a single column.  Moreover pg_crypto
will impose a single key on the column.

Please share your thoughts in which direction i can start analysing this
area for encryption of data specific to a tenant.

The decision to have all tenants in a single database seems rather
unusual to me. Isolating one tenant from adversely impacting another
would seem complicated and I'm not sure how you would implement a clear
security model. Your model has effectively bypassed all the provided PG
facilities for isolation of data. Disaster recovery and business
continuity planning under this model must be a nightmare!

I doubt you can adopt a solution which is solely within the database.
How would the database know which key to use for which rows of data? How
would you select the data for your tenant views if all that data is
encrypted with different keys? How would you manage these keys in a
secure manner?

With the model you have adopted, I would be looking at performing
encryption/decryption at the client level. However, depending on your
data types, this could be challenging. this is really a requirement
which should have been factored into the initial architecture design.
Anything you try to bolt on now is likely to be complex and have
significant performance impact and that is assuming you can re-interpret
the requirement to make the objective feasible.

Yeah, I lost that same arguement in ~2007, where the forces against my push for separation was shouted down with rants on scheme maintenance (divergence) and multiple rollouts per update.  I hadn’t had any coffee before the 9:00am meeting so the hotshot from Amazon got his way.  Then we tried “veils” (a concoction of view and rule re-writing) and we all know how that went.  The company folded before our “next gen” software saw the light of day.

I get the feeling multi-tenancy is, if not the rule these days, at least quite common (on the last of “big iron”?) but it still doesn’t sit well with me.