Skip to content

HOW_INVERTED_INDEX_WORKS.md

Latest commit

 

History

History
251 lines (208 loc) · 12.7 KB

HOW_INVERTED_INDEX_WORKS.md

File metadata and controls

251 lines (208 loc) · 12.7 KB

How the Inverted Index Works

This is how the search index for scdb works under the hood. Keep in mind that scdb is basically an inverted index implemented as a hashmap, backed by a list of doubly-linked cyclic lists of offsets.

Inverted Index

In the most basic of terms, look at the inverted index as a mapping where the keys are the character sets, and the values are the main db offsets of the keys that have those character sets.

Note: In actual sense, the values themselves have some important meta information like expiry, is_deleted etc.

For example:

Consider the following keys: "foo", "fore", "bar", "band", "pig".

Assume that the offsets of these keys in the scdb file are:

"foo": 1, "fore": 2, "bar": 3, "band": 4, "pig": 5

The inverted index would be:

package example
  
_ = map[string][]int{
	"f": [1, 2],
	"b": [3, 4],
	"p": [5],
	"fo": [1, 2],
	"ba": [3, 4],
	"pi": [5],
	"foo": [1],
	"for": [2],
	"bar": [3],
	"ban": [4],
	"pig": [5],
}

Operations

There are five main operations

1. Initialization

  • This creates the search index file if it does not exist
    • adds the 100-byte header basing on user configuration and default values
    • adds the placeholders for the index blocks, each item pre-initialized with a zero.
  • And loads the derived and non-derived properties like 'max_keys', 'max_index_key_len', 'block_size', ' redundant_blocks', 'number_of_index_blocks' ((max_keys / number_of_items_per_index_block).ceil() + redundant_blocks), 'number_of_items_per_index_block' ((block_size / 8).floor()), 'values_start_point' (100 + (net_block_size * number_of_index_blocks)), 'net_block_size' (number_of_items_per_index_block * 8)

2. Add

  1. Get the prefix of the key passed, upto n characters. At the start n = 1.
  2. The prefix supplied is run through a hash function with modulo number_of_items_per_index_block and answer multiplied by 8 to get the byte offset. Let the hashed value be hash
  3. Set index_block_offset to zero to start from the first block.
  4. The index_address is set to index_block_offset + 100 + hash.
  5. The 8-byte offset at the index_address offset is read. This is the first possible pointer to the value entry. Let's call it root_value_offset.
  6. If this root_value_offset is zero, this means that no value has been set for that inverted index key i.e. prefix yet.
    • set the value's is_root to true
    • set the value's next_offset to last_offset
    • set the value's previous_offset to last_offset (i.e. previous offset = next offset = offset of current value)
    • so the value entry (with all its data including index_key_size, expiry (got from ttl from user), kv_address , deleted) is appended to the end of the file at offset last_offset
    • the last_offset is then inserted at index_address in place of the zero
    • the last_offset header is then updated to last_offset + get_size_of_v_entry(v) [get_size_of_v gets the total size of the entry in bytes]
  7. Else if this root_value_offset is non-zero, the value for that prefix has already been set.
    • set the value_offset of the current value to root_value_offset
    • read the value at the value_offset. This is the current value.
    • retrieve the index_key of this current value.
    • if the index_key is the same as the prefix passed:
      • i. If the db_key of the current value entry is equal to the key that is to be added:
        • set the new data from the db into the current value entry i.e. new expiry and the new db_offset.
        • increment n by 1
          • if n is greater than max_index_key_len, stop the iteration and exit.
          • else go back to step 1 ii. Else if the db_key is not equal to the key being added
        • if the next_offset is equal to the root_value_offset, append the new value to the end of this list i.e.
          • Append the value entry (with all its data including index_key_size, expiry (got from ttl from user) deleted, kv_address) is to the end of the file at offset last_offset
          • Set the next_offset of the current value entry (not the newly appended one) to last_offset
          • Set the previous_offset of the newly appended entry to the value_offset
          • Set the next_offset of the newly appended entry to the root_value_offset.
          • the last_offset header is then updated to last_offset + get_size_of_v(v)
        • else
          • set value_offset to next_offset.
          • read the value at value_offset and this becomes the current value.
          • Go back to step (i) - i.e. check db_key against the key passed.
    • else increment the index_block_offset by net_block_size
      • if the new index_block_offset is equal to or greater than the values_start_point, raise the CollisionSaturatedError error. We have run out of blocks without getting a free slot to add the value entry.
      • else go back to step 4.

Note: this uses a form of separate chaining to handle hash collisions. Having multiple index blocks is a form of separate chaining

Performance

  • Time complexity: This operation is O(km+n) where:
    • k = number_of_index_blocks
    • m = max_index_key_len
    • n = length of the longest linked list of values accessed.
  • Auxiliary Space: This operation is O(km+n) where:
    • n = length of the longest db_key in the linked list of values accessed.
    • m = max_index_key_len.
    • k = number_of_index_blocks. If the hash for the given prefix already has offsets associated with it, each will be allocated in memory thus km.

Caveats

  • There is a possibility that when one sets a given index_key or prefix, we may find all index blocks for the given hash already filled. We thus have to throw a CollisionSaturatedError and abort inserting the key. This means that the occurrence of such errors will increase in frequency as the number of keys comes closer to the max_keys value. One possible remedy to this is to add a more redundant index block(s) i.e. increase redundant_blocks. Keep in mind that this consumes extra disk and memory space.

3. Remove

  1. Get the prefix of the key passed, upto n characters. At the start n = 1.
  2. The prefix is run through a hash function with modulo number_of_items_per_index_block and answer multiplied by 8 to get the byte offset. Let the hashed value be hash.
  3. Set index_block_offset to zero to start from the first block.
  4. The index_address is set to index_block_offset + 100 + hash.
  5. The 8-byte offset at the index_address offset is read. This is the first possible pointer to the key-value entry. Let's call it root_value_offset.
  6. Set the value_offset of the current value to root_value_offset
  7. If this root_value_offset is non-zero, it is possible that the value for that key exists.

    • read the value at the value_offset. This is the current_value.

    • retrieve the index_key of this current_value.

    • if the index_key is the same as the prefix:

      • i. retrieve the db_key of this current_value.
      • ii. if the db_key is the same as the key passed:
        • update the is_deleted of the current_value to true
        • if previous_offset equals value_offset:
          • set the previous_value to current_value
        • else:
          • read the value at previous_offset of the current_value. Set that previous_value to that value.
        • if next_offset equals value_offset:
          • set next_value to current_value
        • else if next_offset equals previous_offset
          • set next_value to previous_value
        • else:
        • read the value at next_offset of the current_value. Set that next_value to that value
      • update the next_offset of the previous_value to next_offset of the current_value.
        • update the previous_offset of the next_value to previous_offset of the current_value.
        • if is_root is true for current_value:
          • set the is_root of the next_value to true.
          • set the value at index_address to next_offset
        • if value_offset equals root_value_offset
          • set the value at index_address to 0 i.e. reset it
        • increment n by 1
        • if n is greater than max_index_key_len, exit
        • else go back to step 1

      -iii. else if db_key is not equal to the key passed:

      • if next_offset equals root_value_offset:
      • increment n by 1:
      • if n is greater than max_index_key_len, exit - else go back to step 1 - else:
      • set the value_offset to next_offset
      • read the value at the value_offset. This is the current_value. - go back to step (i)

    • else increment the index_block_offset by net_block_size

      • if the new index_block_offset is equal to or greater than the values_start_point, exit.
      • else go back to step 4.

Performance

  • Time complexity: This operation is O(km+n) where:
    • k = number_of_index_blocks
    • m = max_index_key_len
    • n = length of the longest linked list of values accessed.
  • Auxiliary space: This operation is O(km+n) where:
    • k = number_of_index_blocks
    • m = max_index_key_len
    • n = length of the longest db_key in the linked list of values accessed.

4. Search

  1. Get the lower of the two values: length of the search_term and the max_index_key_len. Let it be n.
  2. Get the prefix i.e. the first n characters/runes of the search_term
  3. The prefix is run through a hash function with modulo number_of_items_per_index_block and answer multiplied by 8 to get the byte offset. Let the hashed value be hash.
  4. Set index_block_offset to zero to start from the first block.
  5. The index_address is set to index_block_offset + 100 + hash.
  6. The 8-byte offset at the index_address offset is read. This is the first possible pointer to the key-value entry. Let's call it root_value_offset.
  7. If this root_value_offset is non-zero, it is possible that the value for that prefix exists.
    • retrieve the value at the root_value_offset. Let it be current_value.
    • retrieve the index_key of the current_value
    • if index_key equals prefix:
      • i. let value_offset equal to root_value_offset
      • ii. retrieve the db_key of the current_value
      • iii. if db_key contains the search_term:
        • add its kv_address to the list of matched_addresses
      • iv. set the value_offset to next_offset of the current_value
      • v. if the current_offset equals root_value_offset
        • read the key, values of the matched_addresses from the main database file and return them - exit
      • vi. else go back to step (ii).
    • else increment the index_block_offset by net_block_size
      • if the new index_block_offset is equal to or greater than the key_values_start_point, stop and return an empty list. No matches found.
      • else go back to step 5.
  8. Else return an empty list. No matches found.

Performance

  • Time complexity: This operation is O(kn) where:
    • k = number_of_index_blocks.
    • n = length of the longest linked list of values accessed.
  • Auxiliary space: This operation is O(kn) where:
    • k = number_of_index_blocks
    • n = length of the longest db_key in the linked list of values accessed.

5. Clear

Clear the entire database.

  1. Initialize a new header basing on the old settings
  2. Write the new header into the file
  3. Shrink the file to the size of the header
  4. Expand the file to the expected size of file if it had headers and index blocks. This fills any gaps with 0s.

Performance

  • Time complexity: This operation is O(1)
  • Auxiliary space: This operation is O(1).

Integration With DB File

  • The inverted index works separately from the BufferPool only until the point of compaction.
  • When the DB file is being compacted, addresses are being moved around and so that requires the index to be rebuilt.
  • This makes compaction an even more expensive operation than it was before.

Optimizations

Pagination

  • skip and limit parameters are provided where:
    • skip is the number of matching records to skip before starting to return. Default is zero.
    • limit is the maximum number of records to return at any one time. Default is zero. When limit is zero, all matched records are returned since it would make no sense for someone to search for zero items.