On Fri, May 5, 2017 at 4:54 PM, Dan Burkert <dan@cloudera.com> wrote:

On Tue, May 2, 2017 at 8:38 PM, Franco Venturi <fventuri@comcast.net> wrote:

first of all thanks for reading through my long post and providing your comments and advice.

You are 100% correct on the TDE column encryption in Oracle; I looked it up again in the 'Introduction to Transparent Data Encryption' in the 'Data Advanced Security Guide' (https://docs.oracle.com/database/121/ASOAG/asotrans.htm#ASOAG10117) and Figure 2-1 clearly shows the keys being stored in the database.
With this piece of information, it doesn't seem to me that Oracle column TDE offers much protection in case of an active attacker who has full access to the the DB server, since there must be a proces somewhere where the database engine is able to retrieve the decryption key for a given column.

Yes, but this could be in a hardware HSM.

 Another interesting piece of information in that chapter is this sentence:

                TDE tablespace encryption also allows index range scans on data in encrypted tablespaces. This is not possible with TDE column encryption.

which makes me think that TDE column encryption must encrypt the data before placing it into the Btree, and therefore is not able to use the Btree for range searches.

That's my interpretation as well.

I think the main reason why an organization would want one or the other type of encryption (client-side vs server-side) is what kind of possible attack they are trying to prevent (and the criteria are often dictated by internal security policies):
        - with server-side encryption, the encrypted data is protected against a disk being lost (the so called 'encryption at rest'), but it is not protected against an active attacker on the server with full access (they could retrieve the key and then decrypt the data).
        - with client-side encryption, the server has no way to decrypt the data and therefore even the active attacker above wouldn't be able to do much with the encrypted data. As I mentioned in my previous post, this is similar to what HDFS does for transparent data encryption and I think it's one of their selling points ('not even root can decrypt the data on HDFS'), and for some IT security groups this may sound attractive.

Root privileges on a machine doesn't necessary guarantee access to the key; the key could be stored remotely, or even on an HSM.

100% agree with your performance concerns that client-side encryption raises (no range scans on the encryped columns, no compression, RLE, etc), to the point that last night I wondered if other people have asked themselves similar questions, and I did find a couple of interesting approaches:
        - CryptDB (http://css.csail.mit.edu/cryptdb/ - the main paper is here: http://people.csail.mit.edu/nickolai/papers/raluca-cryptdb.pdf)
        - ZeroDB (https://opensource.zerodb.com/)

in order to be able to do range scans, for instance CryptDB uses this 'Order Preserving Encryption', which in theory allows to encrypt data in a way that preservers ordering, i.e. Enc(x) < Enc(y) iff x < y; however several research papers after that show that this Order Preserving Encryption leaks a significant amount of information on the encrypted data and is susceptible to frequency and other kind of attacks. As you can imagine there's a lot of academic research actively being done in this field and, even if not ready for prime time, I though I would share these findings.

That's really interesting.  Pretty different threat model being assumed by ZeroDB :).

After this long digression (hopefully not too boring), I agree that the way forward would be to start with looking into the encryption of the file store (I think they are called 'cfiles'; I saw also mentions to some 'delta' files, and I am not sure if they are written the same way and should be encrypted too), and after that the WALs.

Yah, I think cfiles are a good place to start.  AFAIK delta files reuse the cfile machinery when writing to disk. I originally considered recommending looking at the filesystem block manager, but we often do offset lookups into the FS blocks, which I don't think could be supported with encryption.

I think it could be -- if you use CTR mode for encryption, you can support random access, right?

However, I do think it makes sense to consider column-level encryption keys/policies in which case it may be easier to do at a higher level. Though, it may be possible for the higher level to just pass down a key ID into the FS layer when writing a file, so that the policy can be set at a high level while the implementation is done at a lower one.
- Dan

From: "Dan Burkert" <danburkert@apache.org>
To: user@kudu.apache.org
Sent: Tuesday, May 2, 2017 2:54:26 PM

Subject: Re: Data encryption in Kudu

Hi Franco,

Thanks for the writeup!  I'm not an Oracle expert, but my interpretation of the TDE column level encryption documentation/implementation is very different than yours.  As far as I can tell, in both the per-column and table-space encryption modes, encryption/decryption is handled entirely on the Oracle server.  The difference is that column-level encryption will encrypt individual cells on disk (leaving the overall tree/index structure unencrypted), while table-space level encryption will encrypt at the block or file level.

I agree with everything you wrote about the tradoffs involved with client vs server encryption, but I think you are underestimating both the complexity involved with client-side encryption, as well as the performance hit that it would impose.  The loss on encoding, compression, and range predicate pushdown would absolutely kill performance for many important usecases.  The implementation would also be significantly _more_ difficult than server side encryption, because the client would need to manage the encryption keys, encrypt/decrypt data, and the solution would need to be implemented for every client library (of which there are currently two).

For those reasons, I think server side encryption is the way to go with Kudu.  I think you're right that it would slot in as an additional step in the encode -> compress -> encrypt pipeline for blocks.  Because blocks are relatively large (typically > 1 MiB), the overhead of a 16 byte salt and additional MAC are negligible, so we wouldn't need to force the user to make that tradeoff.  Basically, we could get all of the advantages that Oracle's tablespace level encryption provides, but on a per-column basis.  There are a couple of additional complications - we also have a WAL that lives outside of our file block abstraction, and we would almost certainly need to provide encryption for that as well (but perhaps it could be a second step in the process).

In-line responses to some other comments below.

On Sat, Apr 29, 2017 at 8:35 PM, Franco Venturi <fventuri@comcast.net> wrote:

- also from the security point of view, since the encryption happens at the client side, the data that is transfered on the network between the client and the server is already encrypted and there's no need (at least from this point of view) to add a layer of encryption between client and server

I'm skeptical of this.  For instances, every scan request includes the names and types of the columns that the client wishes to scan, and that would be in plaintext without wire encryption.  That would be an issue for some usecases.

- from the security point of view, an attacker with full access to the server would probably be able to decrypt the encrypted data

Could you elaborate on this?  As long as we use an external keystore and intermediate keys, I don't know how an attacker with access to the on-disk files could decrypt them.

- also from a security point of view the server returns the data back in plaintext format; if the data transferred over the network contains sensitive information, it would need an extra encryption layer like TLS or something like that

Correct, and Kudu 1.3 includes TLS wire encryption for exactly this reason.

- as per performance implications, if the encryption on the server side uses something like AES192 or AES256, there are libraries like libcrypto that take advantage of the hardware acceleration for AES encryption on many modern CPUs and therefore I suspect the performance overhead would be limited; this is also indicated by what the Oracle documentation says regarding processing overhead in the case of tablespace encryption in TDE

I agree, I think the overhead of per-block encryption would be pretty minimal.

- it would also require a way to have the server manage these column encryption keys (possibly though additional client API's); I haven't looked yet at the way Oracle handles encryption/decryption keys for the tablespace encryption TDE, but it's on my 'to-do' list

Yah, the normal thing to do here is call out to an external keystore that holds a master encryption key.
- Dan

From: fventuri@comcast.net
To: user@kudu.apache.org
Sent: Wednesday, April 26, 2017 9:48:07 PM

Subject: Re: Data encryption in Kudu

David, Dan, Todd,
thanks for your prompt replies.

At this stage I am just exploring what it would take to implement some sort of data encryption in Kudu.

After reading your comments here are some further thoughts:

- according to the first sentence in this paragraph in the Kudu docs (https://kudu.apache.org/docs/schema_design.html#compression):

         Kudu allows per-column compression using the LZ4Snappy, or zlib compression codecs.

it should be possible to perform per-column encryption by adding 'encryption codecs' right after the compression codecs. I browsed through the code quickly and I think this done when reading/writing a 'cfile' (please correct me if I am wrong). If this is correct, this change could be 'minimally invasive' (at least for the 'cfile' part) and would not require a major overhaul of the Kudu architecture.

- as per the key management aspect, I am not a security expert at all, so I am not sure what would be the best approach here - my thought here is that in most places Kudu is deployed together with HDFS, so it would be 'desirable' if the key management were consistent between the two services; on the other hand, I also realize that the basic premises are fundamentally different: HDFS encrypts everything at the client level and therefore the HDFS engine itself is almost completely unaware that the data it stores is actually encrypted (except for a special file hidden attribute, if I understand correctly), while in Kudu the storage engine must have both the 'public' key (when encrypting) and the 'private' key (when decrypting) otherwise it can't take advantage of knowing the 'structure' of the data (for instance the Bloom filters wouldn't probably work with the key being encrypted). This means for instance that an attacker who is able to gain access to the Kudu tablet servers would probably be able to decrypt the data. Also one way to achieve something similar to what HDFS does (i.e. client-based encryption and data encrypted in-flight) could be perhaps using a one-time client certificate generated by the KMS server, but this would also require changes to the client code.


From: "Todd Lipcon" <todd@cloudera.com>
To: user@kudu.apache.org
Sent: Tuesday, April 25, 2017 3:49:50 PM
Subject: Re: Data encryption in Kudu

Agreed with what Dan said.

I think there are a number of interesting design alternatives to be considered, so before coding it would be great to work through a design document to explore the alternatives. For example, we could try to apply encryption at the 'fs/' layer, which would cover all non-WAL data, but then we would lose the ability to specify encryption on a per-column basis. There are other requirements that need to be ironed out about whether we'd need to support separate encryption keys per column/table/server/etc, whether metadata also needs to be encrypted, etc.


On Tue, Apr 25, 2017 at 10:38 AM, Dan Burkert <danburkert@apache.org> wrote:
Hi Franco,

I think you are right that a client-based approach wouldn't work, because we wouldn't want to encrypt at the level of individual cell values.  That would get in the way of encoding, compression, predicate evaluation, etc.  As you note, adding encryption at the block layer is probably the way to go.  Key management is definitely the tricky issue.  We do have one advantage over HDFS - because Kudu does logical replication, the encryption key can be scoped to a particular tablet server or tablet replica, it wouldn't need to be shared among all replicas.  I haven't done enough research to know if this makes it fundamentally easier to do key management.  I would assume at a minimum we would want to integrate with key providers such an HSM.  It would be good to have a thorough review of existing solutions in the space, such as TDE and the Hadoop KMS.  Is this something you are interested in working on?

- Dan

On Tue, Apr 25, 2017 at 8:30 AM, David Alves <davidralves@gmail.com> wrote:
Hi Franco

  Dan, Alexey, Todd are our security experts.
  Folks, thoughts on this?


On Mon, Apr 24, 2017 at 7:08 PM, <fventuri@comcast.net> wrote:
Over the weekend I started looking at what it would take to add data encryption to Kudu (besides using filesystem encryption via dm-crypt or something like that).

Here are a few notes - please feel free to comment on them and add suggestions:

- reading through this mailing list, it looks like this feature has been asked a couple of times but last year, but from what I can tell, noone is currently working on it.
- a client-based approach to encryption like the one used by HDFS wouldn't work (at least out of the box) because for instance encrypting the primary key at the client would prevent being able to have range filters for scans; it might work for the columns that are not part of the primary key
- there's already code in Kudu for several compression codecs (LZ4, gzip, etc); I thought it would be possible to add similar code for encryption codecs (to be applied after the compression, of course)
- the WAL log files and delta files should be similarly encrypted too
- not sure what would be the best way to manage the key - I see that in HDFS they use a double key mechanism, where the encryption key for the data file is itself encrypted with the allowed user key and this whole process is managed by an external Key Management Service

Thanks in advance for your ideas and suggestions,

Todd Lipcon
Software Engineer, Cloudera

Todd Lipcon
Software Engineer, Cloudera