Add documentation for fuzzy worker.

diff --git a/doc/markdown/workers/fuzzy_storage.md b/doc/markdown/workers/fuzzy_storage.md
index e69de29bb2d1d6434b8b29ae775ad8c2e48c5391..0aed911fc6bfef0614df475f4f76206fdf5f113d 100644 (file)
--- a/doc/markdown/workers/fuzzy_storage.md
+++ b/doc/markdown/workers/fuzzy_storage.md
@@ -0,0 +1,128 @@
+# Fuzzy storage worker
+
+Fuzzy storage worker is intended to store fuzzy hashes of messages.
+
+## Protocol format
+
+Fuzzy storage accepts requests using `UDP` protocol with the following structure:
+
+~~~C
+struct fuzzy_cmd  { /* attribute(packed) */
+       unit8_t version;        /* command version, must be 0x2 */
+       unit8_t cmd;            /* numeric command */
+       unit8_t shingles_count; /* number of shingles */
+       unit8_t flag;           /* flag number */
+       int32_t value;          /* value to store */
+       uint32_t tag;           /* random tag */
+       char digest[64];        /* blake2b digest */
+};
+~~~
+
+All numbers are in host byte order, so if you want to check fuzzy hashes from a
+host with different byte order you need some additional conversions (not currently
+supported by rspamd). In future, rspamd might use little endian byte order for all
+operations.
+
+Fuzzy storage accepts the following commands:
+- `FUZZY_CHECK` - check for a fuzzy hash
+- `FUZZY_ADD` - add a new hash
+- `FUZZY_DEL` - remove a hash
+
+`flag` field is used to store different hashes in a single storage. For example,
+it allows to store blacklists and whitelists in the same fuzzy storage worker. 
+A client should set the `flag` field when adding or deleting hashes and check it
+when querying for a hash.
+
+`value` is added to the currently stored value of a hash if that hash has been found.
+This field can handle negative numbers as well.
+
+`tag` is used to distinguish requests by a client. Fuzzy storage just sets this
+field in the reply equal to the value in the request.
+
+`digest` field contains the content of hash. Currently, rspamd uses `blake2b` hash
+in its binary form granting the `2^512` of possible hashes with negligible collisions
+probability. At the same time, rspamd saves the legacy format of fuzzy hashes by
+means of this field. Old rspamd can work with legacy hashes only.
+
+`shingles_count` defines how many `shingles` are attached to this command.
+Currently, rspamd uses 32 shingles and this value thus should be 32 for commands
+with shingles. Shingles should be included in the same packet and follow the command as
+an array of int64_t values. Please note, that rspamd rejects commands that have wrong
+shingles count or their size is not equal to the desired one:
+
+       sizeof(fuzzy_cmd) + shingles_count * sizeof(int64_t)
+       
+Reply format of fuzzy storage is also presented as a structure:
+
+~~~C
+struct fuzzy_cmd  { /* attribute(packed) */
+       int32_t value;
+       uint32_t flag;
+       uint32_t tag;
+       float prob;
+};
+~~~
+
+`prob` field is used to store the probability of match. This value is changed from
+`0.0` (no match) to `1.0` (full match).
+
+## Storage format
+
+Rspamd fuzzy storage uses `sqlite3` for storing hashes. All update operations are
+performed in a transaction which is committed to the main database approximately once
+per minute. `VACUUM` command is executed on startup and hashes expiration is performed
+at the termination of rspamd fuzzy storage worker.
+
+Here is the internal database structure:
+
+```
+CREATE TABLE digests(id INTEGER PRIMARY KEY,
+       flag INTEGER NOT NULL,
+       digest TEXT NOT NULL,
+       value INTEGER,
+       time INTEGER);
+       
+CREATE TABLE shingles(value INTEGER NOT NULL,
+       number INTEGER NOT NULL,
+       digest_id INTEGER REFERENCES digests(id) ON DELETE CASCADE ON UPDATE CASCADE);
+```
+
+Since rspamd uses normal sqlite3 you can use all tools for working with the hashes
+database to perform, for example backup or analysis.
+
+## Operation notes
+
+To check a hash, rspamd fuzzy storage initially queries for the direct match using
+`digest` field as a key. If that match succeed then the value is returned immediately.
+Otherwise, if a command contains shingles then rspamd checks for fuzzy match trying
+to find each shingle's value. If more than 50% of shingles matches the same digest
+then rspamd returns that digest's value and the probability of match that means
+generally `match_count / shingles_count`.
+
+## Configuration
+
+Fuzzy storage accepts the following extra options:
+
+- `database` - path to the sqlite storage
+- `expire` - time value for hashes expiration
+- `allow_map` - string, array of strings or a map of IP addresses that are allowed
+to perform changes to fuzzy storage
+
+Here is an example configuration of fuzzy storage:
+
+~~~nginx
+worker {
+   type = "fuzzy";
+   bind_socket = "*:11335";
+   hash_file = "${DBDIR}/fuzzy.db"
+   expire = 90d;
+   allow_update = "127.0.0.1";
+}
+~~~
+
+## Compatibility notes
+
+Rspamd fuzzy storage of version `0.8` can work with rspamd clients of all versions,
+however, all updates from legacy versions (less that `0.8`) won't update fuzzy shingles
+database. Rspamd [fuzzy check module](../modules/fuzzy_check.md) can work **only**
+with the recent rspamd fuzzy storage (it won't get anything from the legacy storages).
\ No newline at end of file