summaryrefslogtreecommitdiff
path: root/example.py
diff options
context:
space:
mode:
authorGreg Bowyer <gbowyer@fastmail.co.uk>2014-02-19 14:42:40 -0800
committerGreg Bowyer <gbowyer@fastmail.co.uk>2014-02-19 14:42:40 -0800
commit96c9ce0c4a2ab515c581062117d5a6221f1f2036 (patch)
treea679fc840ae486807207158fb85a483dbd1faef7 /example.py
parent4abf7ee1fbbdc47c8cb7b35f2600e58f1f95e6bb (diff)
downloadkafka-python-96c9ce0c4a2ab515c581062117d5a6221f1f2036.tar.gz
Make it possible to read and write xerial snappy
Fixes mumrah/kafka-python#126 TL;DR ===== This makes it possible to read and write snappy compressed streams that are compatible with the java and scala kafka clients (the xerial blocking format)) Xerial Details ============== Kafka supports transparent compression of data (both in transit and at rest) of messages, one of the allowable compression algorithms is Google's snappy, an algorithm which has excellent performance at the cost of efficiency. The specific implementation of snappy used in kafka is the xerial-snappy implementation, this is a readily available java library for snappy. As part of this implementation, there is a specialised blocking format that is somewhat none standard in the snappy world. Xerial Format ------------- The blocking mode of the xerial snappy library is fairly simple, using a magic header to identify itself and then a size + block scheme, unless otherwise noted all items in xerials blocking format are assumed to be big-endian. A block size (```xerial_blocksize``` in implementation) controls how frequent the blocking occurs 32k is the default in the xerial library, this blocking controls the size of the uncompressed chunks that will be fed to snappy to be compressed. The format winds up being | Header | Block1 len | Block1 data | Blockn len | Blockn data | | ----------- | ---------- | ------------ | ---------- | ------------ | | 16 bytes | BE int32 | snappy bytes | BE int32 | snappy bytes | It is important to not that the blocksize is the amount of uncompressed data presented to snappy at each block, whereas the blocklen is the number of bytes that will be present in the stream, that is the length will always be <= blocksize. Xerial blocking header ---------------------- Marker | Magic String | Null / Pad | Version | Compat ------ | ------------ | ---------- | -------- | -------- byte | c-string | byte | int32 | int32 ------ | ------------ | ---------- | -------- | -------- -126 | 'SNAPPY' | \0 | variable | variable The pad appears to be to ensure that SNAPPY is a valid cstring, and to align the header on a word boundary. The version is the version of this format as written by xerial, in the wild this is currently 1 as such we only support v1. Compat is there to claim the minimum supported version that can read a xerial block stream, presently in the wild this is 1. Implementation specific details =============================== The implementation presented here follows the Xerial implementation as of its v1 blocking format, no attempts are made to check for future versions. Since none-xerial aware clients might have persisted snappy compressed messages to kafka brokers we allow clients to turn on xerial compatibility for message sending, and perform header sniffing to detect xerial vs plain snappy payloads.
Diffstat (limited to 'example.py')
0 files changed, 0 insertions, 0 deletions