-- LarsAronsson - 30 Jun 2004

These notes were written on 30 June 2004, and doesn't handle the config files or the web db.

Nutch file formats from the bottom up

Nutch version 0.5

-- Differences noted by MattKangas - 04 Jan 2005

A segment now consists of five subdirectories, each containing an ArrayFile:

Subdirectory Value datatype Variable
fetchlist net.nutch.pagedb.FetchListEntry fetchList
fetcher net.nutch.fetcher.FetcherOutput fetcherWriter
content net.nutch.protocol.Content contentWriter
parse_text net.nutch.parse.ParseText parseTextWriter
parse_data net.nutch.parse.ParseData parseDataWriter

FetcherOutput is changed:

  1 byte version (value 4, was 3)
  FetchListEntry as specified above
  16 bytes MD5 hash
  1 byte status
  8 bytes (long) Java milliseconds fetchdate

New class: net.nutch.protocol.Content

  1 byte version (value 1)
  UTF8 string url
  UTF8 string base
  compressed byte array content
  UTF8 string contentType
  java.util.Properties metadata

New class: net.nutch.parse.ParseText

  1 byte version (value 1)
  compressed byte array text

New class: net.nutch.parse.ParseData

  1 byte version (value 1)
  UTF8 string title
  4 bytes integer totalOutlinks
  a list of net.nutch.fetcher.Outlink objects:
          UTF8 string URL
          UTF8 string anchor
  java.util.Properties metadata

Nutch version 0.4

Nutch 0.4 was released on May 25, 2004 (the previous version 0.3 was from June 17, 2003). The Java source code consists of 165 files comprising 37,178 lines of code.

Nutch implements its own serialization to store serialized Java data types and structures on file. The interface net.nutch.io.Writable must be implemented for all such data types. In some cases, long text strings are stored in GZIP (Gnu ZIP) compressed format.

The abstract class nutch.io.VersionedWritable prepends a byte indicating the version of the data structure, typically \001.

Nutch uses Java's native UTF-8 character set, and the class net.nutch.io.UTF8 for writing short strings to files. The UTF8 class limits the length of strings to 0xffff/3 or 21845 bytes. The function UTF8.write() uses java.io.DataOutput.writeShort() to prepend the length of the string. This is why the two bytes \000\003 is seen before a three letter word in a file. The zero byte is thus not a null termination of the previous string (strings are not null terminated), but the most significant byte of the 16 bit short integer indicating the length of the following string.

Nutch relies heavily on mappings (associative arrays) from keys to values. The class net.nutch.io.SequenceFile is a flat file of keys and values. The first four bytes of each such file are ASCII "SEQ" and \001 (C-a), followed by the Java class names of keys and values, written as UTF8 strings, e.g. "SEQ\001\000\004long\000\004long", for a mapping from long integers to long integers. After that follows the key-value pairs. Each pair is introduced by four bytes telling the length in bytes of the pair (excluding the eight length bytes) and four bytes telling the length of the key. The typical long (64 bit) integer is 8 bytes and a long-to-long mapping will have pairs of length 16 bytes, e.g.

  00 00 00 10                                   int length of pair = 0x10 = 16 bytes
  00 00 00 08                                   int length of key  = 0x08 =  8 bytes
  00 00 00 00 00 00 02 80       long key = 0x280 = 640
  00 00 00 00 00 0a 42 9b       long value = 0xa429b = 672411

To economize the handling of large data volumes, net.nutch.io.MapFile manages a mapping as two separate files in a subdirectory of its own. The large "data" file stores all keys and values, sorted by the key. The much smaller "index" file points to byte offsets in the data file for a small sample of keys. Only the index file is read into memory.

net.nutch.io.ArrayFile is a specialization of MapFile where the keys are long integers.

The Java files in net.nutch.io.* comprise 2556 lines of source code. The biggest one is Sequencefile.java, which contains a Writer (112 lines), a Reader (138 lines), a BufferedRandomAccessFile (140 lines) and a Sorter (389 lines).

When Nutch crawls the web, each resulting segment has four subdirectories, each containing an ArrayFile (a MapFile having keys that are long integers):

Subdirectory Value datatype Variable
fetchlist net.nutch.pagedb.FetchListEntry fetchList
fetcher net.nutch.fetcher.FetcherOutput fetcherDb
fetcher_content net.nutch.fetcher.FetcherContent rawDb
fetcher_text net.nutch.fetcher.FetcherText strippedDb

Crawling is performed by net.nutch.fetcher.Fetcher which starts a number of parallel FetcherThread?. Each thread gets an URL from the fetchList, checks robots.txt, retrieves the contents and appends the results to fetcherDb, rawDb, and strippedDb.

The FetchListEntry is represented thus:

  1 byte version (value should be 2),
  1 byte flag (value 1 = true if page should be fetched)
  page, as defined by net.nutch.db.Page:
         1 byte version (value should be 4)
         UTF8 string URL (primary key)
         16 bytes (128 bit) MD5 hash of contents
         8 bytes (64 bit long) Java milliseconds when page should be refetched
         1 byte number of failed attempts
         1 byte fetch interval in days
         4 bytes Java float score
         4 bytes Java float next score
  4 bytes number of anchors
  a list of anchors represented as UTF8 strings

The FetcherOutput is all of the fetcher's output except the raw and stripped versions of the contents:

  1 byte version (value 3)
  FetchListEntry as specified above
  16 bytes MD5 hash
  1 byte status
  UTF8 string title
  4 bytes integer totalOutlinks
  a list of net.nutch.fetcher.Outlink objects:
          UTF8 string URL
          UTF8 string anchor
  8 bytes (long) Java milliseconds fetchdate

The FetcherContent is the raw contents stored in GZIP:

  1 byte version (value 1)
  compressed byte array

The FetcherText is the text conversion of page's content, stored in GZIP:

  1 byte version (value 1)
  compressed byte array

NutchFileFormats (last edited 2009-09-20 23:09:42 by localhost)