We would like to contribute a codec that enables the encryption of sensitive data in the index that has been developed as part of an engagement with a customer. We think that this could be of interest for the community.
Below is a description of the project.
In comparison with approaches where all data is encrypted (e.g., file system encryption, index output / directory encryption), encryption at a codec level enables more fine-grained control on which block of data is encrypted. This is more efficient since less data has to be encrypted. This also gives more flexibility such as the ability to select which field to encrypt.
Some of the requirements for this project were:
- The performance impact of the encryption should be reasonable.
- The user can choose which field to encrypt.
- Key management: During the life cycle of the index, the user can provide a new version of his encryption key. Multiple key versions should co-exist in one index.
- Block tree terms index and dictionary
- Compressed stored fields format
- Compressed term vectors format
- Doc values format (prototype based on an encrypted index output) - this will be submitted as a separated patch
- Index upgrader: command to upgrade all the index segments with the latest key version available.
One index segment is encrypted with a single key version. An index can have multiple segments, each one encrypted using a different key version. The key version for a segment is stored in the segment info.
The provided codec is abstract, and a subclass is responsible in providing an implementation of the cipher factory. The cipher factory is responsible of the creation of a cipher instance based on a given key version.
The encryption model is based on AES/CBC with padding. Initialisation vector (IV) is reused for performance reason, but only on a per format and per segment basis.
While IV reuse is usually considered a bad practice, the CBC mode is somehow resilient to IV reuse. The only "leak" of information that this could lead to is being able to know that two encrypted blocks of data starts with the same prefix. However, it is unlikely that two data blocks in an index segment will start with the same data:
- Stored Fields Format: Each encrypted data block is a compressed block (~4kb) of one or more documents. It is unlikely that two compressed blocks start with the same data prefix.
- Term Vectors: Each encrypted data block is a compressed block (~4kb) of terms and payloads from one or more documents. It is unlikely that two compressed blocks start with the same data prefix.
- Term Dictionary Index: The term dictionary index is encoded and encrypted in one single data block.
- Term Dictionary Data: Each data block of the term dictionary encodes a set of suffixes. It is unlikely to have two dictionary data blocks sharing the same prefix within the same segment.
- DocValues: A DocValues file will be composed of multiple encrypted data blocks. It is unlikely to have two data blocks sharing the same prefix within the same segment (each one will encodes a list of values associated to a field).
To the best of our knowledge, this model should be safe. However, it would be good if someone with security expertise in the community could review and validate it.
We report here a performance benchmark we did on an early prototype based on Lucene 4.x. The benchmark was performed on the Wikipedia dataset where all the fields (id, title, body, date) were encrypted. Only the block tree terms and compressed stored fields format were tested at that time.
The indexing throughput slightly decreased and is roughly 15% less than with the base Lucene.
The merge time slightly increased by 35%.
There was no significant difference in term of index size.
With respect to query throughput, we observed no significant impact on the following queries: Term query, boolean query, phrase query, numeric range query.
We observed the following performance impact for queries that needs to scan a larger portion of the term dictionary:
- prefix query: decrease of ~25%
- wildcard query (e.g., “fu*r”): decrease of ~60%
- fuzzy query (distance 1): decrease of ~40%
- fuzzy query (distance 2): decrease of ~80%
We can see that the decrease of performance is relative to the size of the dictionary scan.
We observed a decrease of performance that is relative to the size of the set of documents to be retrieved:
- ~20% when retrieving a medium set of documents (100)
- ~30/40% when retrieving a large set of documents (1000)
- compressed stored field do not keep order of fields since non-encrypted and encrypted fields are stored in separated blocks.
- the current implementation of the cipher factory does not enforce the use of AES/CBC. We are planning to add this to the final version of the patch.
- the current implementation does not change the IV per segment. We are planning to add this to the final version of the patch.
- the current implementation of compressed stored fields decrypts a full compressed block even if a small portion is decompressed (high impact when storing very small documents). We are planning to add this optimisation to the final version of the patch. The overall document retrieval performance might increase with this optimisation.
The codec has been implemented as a contrib. Given that most of the classes were final, we had to copy most of the original code from the extended formats. At a later stage, we could think of opening some of these classes to extend them properly in order to reduce code duplication and simplify code maintenance.