Doing encryption on the client has the following benefits (most of them already mentioned in the thread):
* brokers don't have additional CPU load * brokers never see the data in unencrypted form (Kafka admins cannot snoop) * secure multi-tenancy (keys are 100% on the client space) * no need to secure Kafka wire transport, client-broker and broker-broker (data is already encrypted) It would be highly desirable, even if encryption is done on the client side, that encryption is 'transparent'. Similar to how HDFS encryption works, it is not the client writing/reading a topic the one that decides to encrypt/decrypt but the broker is the one telling that to the client and providing encrypted encryption keys for the tasks.The client-broker protocol would have to be augmented to carry the encrypted encryption key, plus logic to handle redistribution to existing clients due to key rotation. A nice thing about doing broker side encryption though is that you can shut off clients at any time and they won't see unencrypted data anymore. But this means the brokers will have to deal with the client ACLs for encryption (i'd rather leave that outside of Kafka and being a concern of the KMS system). You could achieve similar functionality on client side encryption, by removing the client from the ACLs in the KMS and doing a key rotation, then the client will not be able to decrypt any messages using the new key (though all previous ones using the key that the client already has will be visible to the client). Thanks. On Mon, Aug 3, 2015 at 10:22 AM, Gwen Shapira <g...@confluent.io> wrote: > If I understand you correctly, you are saying that the kerberos keytab that > the broker uses to authenticate with the KMS will be somewhere on the > broker node and can be used by a malicious admin. > > I agree this is a valid concern. > I am not opposed to client-side encryption, I am more concerned that the > modifications this will require in Kafka broker implementation make the > idea impractical. And obviously, as in any security discussion - there are > lots of details regarding key exchange, management and protection that are > critical. > > Perhaps given a design doc, we can better evaluate the proposed tradeoffs. > > Gwen > > > > On Sat, Aug 1, 2015 at 10:10 AM, Don Bosco Durai <bo...@apache.org> wrote: > > > >Any reason you think its better to let the clients handle it? > > Gwen, I agree with Todd, depending on the goal, the requirements might > > vary. If the goal is that someone stills the disk, then they should be > > able to access the data, then encrypting at Broker is enough. However, if > > the requirement is that the admin/operator should not be able to access > > the data, then client side is the only option. > > > > Hadoop/HDFS transparent data encryption has a similar philosophy, where > > the actual encryption/decryption happens at the client side. > > > > >1. Key management > > Hadoop common has a KMS. And there are industry standards like KMIP. If > > Broker does the encrypt/decrypt, then the solution is much easier. If the > > client does it, then sharing the key would be a challenge. It might be > > even necessary to use asymmetric encryption to limit sharing of the keys. > > > > Bosco > > > > > > > > > > On 7/31/15, 9:31 PM, "Jiangjie Qin" <j...@linkedin.com.INVALID> wrote: > > > > >I agree with Todd, the major concern I have is still the complexity on > > >broker which can kill the performance - which a key advantage of Kafka. > I > > >think there are two separate issues here: > > >1. Key management > > >2. the actual encryption/decryption work. > > > > > >Personally I think it might be OK to have [1] supported in Kafka given > we > > >might need to be compatible with different key management system anyway. > > >But we should just make Kafka compatible with other key management > systems > > >instead of letting Kafka itself to manage the keys. For [2], I think we > > >should keep it on the client side. > > > > > >Jiangjie (Becket) Qin > > > > > >On Fri, Jul 31, 2015 at 5:06 PM, Todd Palino <tpal...@gmail.com> wrote: > > > > > >> 1 - Yes, authorization combined with encryption does get us most of > the > > >>way > > >> there. However, depending on the auditor it might not be good enough. > > >>The > > >> problem is that if you are encrypting at the broker, then by > definition > > >> anyone who has access to the broker (i.e. operations staff) have > access > > >>to > > >> the data. Consider the case where you are passing salary and other > > >> information through the system, and those people do not need a view of > > >>it. > > >> I admit, the 90% solution might be better here than going for a > perfect > > >> solution, but it is something to think about. > > >> > > >> 2 - My worry is people wanting to integrate with different key > systems. > > >>For > > >> example, one person may be fine with providing it in a config file, > > >>while > > >> someone else may want to use the solution from vendor A, someone else > > >>wants > > >> vendor B, and yet another person wants this obscure hardware-based > > >>solution > > >> that exists elsewhere. > > >> > > >> The compaction concern is definitely a good one I hadn't thought of. > I'm > > >> wondering if it's reasonable to just say that compaction will not work > > >> properly with encrypted keys if you do not have consistent encryption > > >>(that > > >> is, the same string encrypts to the same string every time). > > >> > > >> Ultimately I don't like the idea of the broker doing any > encrypt/decrypt > > >> steps OR compression/decompression. This is all CPU overhead that > you're > > >> concentrating in one place instead of distributing the load out to the > > >> clients. Now yes, I know that the broker decompresses to check the CRC > > >>and > > >> assign offsets and then compresses, and we can potentially avoid the > > >> compression step with assigning the batch an offset and a count > instead > > >>but > > >> we still need to consider the CRC. Adding encrypt/decrypt steps adds > > >>even > > >> more overhead and it's going to get very difficult to handle even 2 > > >>Gbits > > >> worth of traffic at that rate. > > >> > > >> There are other situations that concern me, such as revocation of > keys, > > >>and > > >> I'm not sure whether it is better with client-based or server-based > > >> encryption. For example, if I want to revoke a key with client-based > > >> encryption it becomes similar to how we handle Avro schemas > (internally) > > >> now - you change keys, and depending on what your desire is you either > > >> expire out the data for some period of time with the older keys, or > you > > >> just let it sit there and your consuming clients won't have an issue. > > >>With > > >> broker-based encryption, the broker has to work with the multiple keys > > >> per-topic. > > >> > > >> -Todd > > >> > > >> > > >> On Fri, Jul 31, 2015 at 2:38 PM, Gwen Shapira <gshap...@cloudera.com> > > >> wrote: > > >> > > >> > Good points :) > > >> > > > >> > 1) Kafka already (pending commit) has an authorization layer, so > > >> > theoretically we are good for SOX, HIPAA, PCI, etc. Transparent > broker > > >> > encryption will support PCI > > >> > never-let-unencrypted-card-number-hit-disk. > > >> > > > >> > 2) Agree on Key Management being complete PITA. It may better to > > >> > centralize this pain in the broker rather than distributing it to > > >> > clients. Any reason you think its better to let the clients handle > it? > > >> > The way I see it, we'll need to handle key management the way we did > > >> > authorization - give an API for interfacing with existing systems. > > >> > > > >> > More important, we need the broker to be able to decrypt and encrypt > > >> > in order to support compaction (unless we can find a cool > > >> > key-uniqueness-preserving encryption algorithm, but this may not be > as > > >> > secure). I think we also need the broker to be able to re-compress > > >> > data, and since we always encrypt compressed bits (compressing > > >> > encrypted bits doesn't compress), we need the broker to decrypt > before > > >> > re-compressing. > > >> > > > >> > > > >> > > > >> > On Fri, Jul 31, 2015 at 2:27 PM, Todd Palino <tpal...@gmail.com> > > >>wrote: > > >> > > It does limit it to clients that have an implementation for > > >>encryption, > > >> > > however encryption on the client side is better from an auditing > > >>point > > >> of > > >> > > view (whether that is SOX, HIPAA, PCI, or something else). Most of > > >> those > > >> > > types of standards are based around allowing visibility of data to > > >>just > > >> > the > > >> > > people who need it. That includes the admins of the system (who > are > > >> often > > >> > > not the people who use the data). > > >> > > > > >> > > Additionally, key management is a royal pain, and there are lots > of > > >> > > different types of systems that one may want to use. This is a > > >>pretty > > >> big > > >> > > complication for the brokers. > > >> > > > > >> > > -Todd > > >> > > > > >> > > > > >> > > On Fri, Jul 31, 2015 at 2:21 PM, Gwen Shapira > > >><gshap...@cloudera.com> > > >> > wrote: > > >> > > > > >> > >> I've seen interest in HDFS-like "encryption zones" in Kafka. > > >> > >> > > >> > >> This has the advantage of magically encrypting data at rest > > >>regardless > > >> > >> of which client is used as a producer. > > >> > >> Adding it on the client side limits the feature to the java > client. > > >> > >> > > >> > >> Gwen > > >> > >> > > >> > >> On Fri, Jul 31, 2015 at 1:20 PM, eugene miretsky > > >> > >> <eugene.miret...@gmail.com> wrote: > > >> > >> > I think that Hadoop and Cassandra do [1] (Transparent > Encryption) > > >> > >> > > > >> > >> > We're doing [2] (on a side note, for [2] you still need > > >> > authentication on > > >> > >> > the producer side - you don't want an unauthorized user writing > > >> > garbage). > > >> > >> > Right now we have the 'user' doing the encryption and > submitting > > >> raw > > >> > >> bytes > > >> > >> > to the producer. I was suggesting implementing an encryptor in > > >>the > > >> > >> > producer itself - I think it's cleaner and can be reused by > other > > >> > users > > >> > >> > (instead of having to do their own encryption) > > >> > >> > > > >> > >> > Cheers, > > >> > >> > Eugene > > >> > >> > > > >> > >> > On Fri, Jul 31, 2015 at 4:04 PM, Jiangjie Qin > > >> > <j...@linkedin.com.invalid > > >> > >> > > > >> > >> > wrote: > > >> > >> > > > >> > >> >> I think the goal here is to make the actual message stored on > > >> broker > > >> > to > > >> > >> be > > >> > >> >> encrypted, because after we have SSL, the transmission would > be > > >> > >> encrypted. > > >> > >> >> > > >> > >> >> In general there might be tow approaches: > > >> > >> >> 1. Broker do the encryption/decryption > > >> > >> >> 2. Client do the encryption/decryption > > >> > >> >> > > >> > >> >> From performance point of view, I would prefer [2]. It is just > > >>in > > >> > that > > >> > >> >> case, maybe user does not necessarily need to use SSL anymore > > >> because > > >> > >> the > > >> > >> >> data would be encrypted anyway. > > >> > >> >> > > >> > >> >> If we let client do the encryption, there are also two ways to > > >>do > > >> so > > >> > - > > >> > >> >> either we let producer take an encryptor or users can do > > >> > >> >> serialization/encryption outside the producer and send raw > > >>bytes. > > >> The > > >> > >> only > > >> > >> >> difference between the two might be flexibility. For example, > if > > >> > someone > > >> > >> >> wants to know the actual bytes of a message that got sent over > > >>the > > >> > wire, > > >> > >> >> doing it outside the producer would probably more preferable. > > >> > >> >> > > >> > >> >> Jiangjie (Becket) Qin > > >> > >> >> > > >> > >> >> On Thu, Jul 30, 2015 at 12:16 PM, eugene miretsky < > > >> > >> >> eugene.miret...@gmail.com > > >> > >> >> > wrote: > > >> > >> >> > > >> > >> >> > Hi, > > >> > >> >> > > > >> > >> >> > Based on the security wiki page > > >> > >> >> > <https://cwiki.apache.org/confluence/display/KAFKA/Security > > > > >> > >> encryption > > >> > >> >> of > > >> > >> >> > data at rest is out of scope for the time being. However, we > > >>are > > >> > >> >> > implementing encryption in Kafka and would like to see if > > >>there > > >> > is > > >> > >> >> > interest in submitting a patch got it. > > >> > >> >> > > > >> > >> >> > I suppose that one way to implement encryption would be to > > >>add > > >> an > > >> > >> >> > 'encrypted key' field to the Message/MessageSet structures > in > > >> the > > >> > >> >> > wire protocole - however, this is a very big and fundamental > > >> > change. > > >> > >> >> > > > >> > >> >> > A simpler way to add encryption support would be: > > >> > >> >> > 1) Custom Serializer, but it wouldn't be compatible with > other > > >> > custom > > >> > >> >> > serializers (Avro, etc. ) > > >> > >> >> > 2) Add a step in KafkaProducer after serialization to > encrypt > > >> the > > >> > >> data > > >> > >> >> > before it's being submitted to the accumulator (encryption > is > > >> done > > >> > in > > >> > >> the > > >> > >> >> > submitting thread, not in the producer io thread) > > >> > >> >> > > > >> > >> >> > Is there interest in adding #2 to Kafka? > > >> > >> >> > > > >> > >> >> > Cheers, > > >> > >> >> > Eugene > > >> > >> >> > > > >> > >> >> > > >> > >> > > >> > > > >> > > > > > > >
