tarantool / memcached

Memcached protocol 'wrapper' for tarantool
BSD 2-Clause "Simplified" License
28 stars 10 forks source link

Memcached range operation #3

Open bigbes opened 8 years ago

bigbes commented 8 years ago

Overview

Several implementations of servers speaking the memcached protocol have backends that allow for inexpensive operations over ranges of keys. The purpose of this document is to standardize the protocol for making use of these operations.

General Properties of Range Commands

Commands that take ranges all follow the same pattern of key representation. Each range representation will have a minimum and maximum range as well as flags to indicate whether each end is inclusive or exclusive. Either end may be null indicating there's no limit on the side where the null appears. Keys should be treated as byte arrays for range comparisons. The implementation is not expected to perform any type of character conversion or collation.

Abstract Representation of Ranges

In the binary protocol, we've got some extra space and can apply that to defining ranges. The "key" field will be used as the start key. A key length of zero represents null as an empty key can't be stored. The extra info will contain the following info:

+-------------+------+
| Fields      | Size |
+-------------+------+
| End key len | 16   |
| Reserved    | 8    |
| Flags       | 8    |
| Max results | 32   |
+-------------+------+

Flags is a bitmask indicating whether the ends are inclusive or exclusive.

Max results represents the number of objects the operation is permitted to affect. A value of zero indicates there should be no limit.

A server may enforce a maximum limit and respond with an error if the requested limit exceeds the limit enforced by the server (or is zero).

The end key immediately follows the maximum results.

Response

Each requests will generally have many responses and will return similarly to the way stats works in the binary protocol. The final response will be one with an empty key. The binary protocol allows for commands to operate in a "quiet mode" where most responses are quelled. No such facility is available for the text protocol as it can't be done consistently without introducing new verbs (and then can't be done safely).

Text Protocol Generalities

The general form of a range command is as follows (please see details for how this varies by command):

CMD <start inclusion> <end inclusion> <max items> <start key> [end key]\r\n

The end key is optional. If not specified, it is considered null. Null is not explicitly included for the start key as there is a natural minimum of keys which can be used in its place.

The inclusion flags are either 0 or 1. If 0, the endpoint is not included. If 1, the endpoint is included.

This is different from the rget as specified in memcachedb to allow for infinite ranges.

Supported commands

get

Ranged get operations allow multiple key/value pairs returned for a single key request.

set

Ranged set operations allow for a range of existing items to be set to a specific value.

For example, if you have stats stored in a server with keys prefixed as stats. you can reset all of them in a single operation by setting (stats., stats/) to "0". This range specifies everything that begins with "stats." excluding "stats." itself.

Upon batch mutation, each modified value will receive a new distinct CAS identifier.

Binary Protocol Details

Each modified value will respond as a plain set does although the key will be included in the response. The final response will contain no key.

In the case of a quiet request, no response will be given unless the command fails.

Text Protocol Details

The command for text set is rset and has the following form:

rset <start inclusion> <end inclusion> <max items> <flags> <exptime> <bytes> <start key> [end key]\r\n
<data>\r\n

The response will be similar to a gets, but with no values included. Example:

VALUE <key> <flags> 0 [<cas>]\r\n
\r\n

append

Ranged append will append a value to every item within a range.

For both protocols, the response will be in the form of their respective protocol-specific response in ranged set.

Text Protocol Details

The command for text append has the following structure:

rappend <start inclusion> <end inclusion> <max items> <bytes> <start key> [end key]\r\n
<data>\r\n

prepend

Prepend follows the same pattern of append.

delete

Ranged delete will remove all values in the specified range.

For both protocols, the response will be in the form of their respective protocol-specific response in ranged set.

incr

Ranged incr will increment all existing keys within a range by the same amount. This follows the same patterns as existing increment with the obvious exception that it will not initialize keys since it's only modifing things that already exist.

Binary Protocol Details

The extra data includes the standard incr/decr extra data after the ranged extra data.

The responses will include a key and value for every modified item.

Text Protocol Details

The text command for range increment is rincr and is in the following form:

rincr <start inclusion> <end inclusion> <max items> <value> <start key> [end key]\r\n

Responses are the same as a multi-gets.

decr

Ranged decr follows ranged incr.

Command Reference

The following table is a quick reference to all of the commands defined herein. Commands omitted from the text column are intentional as they are not defined outside of the binary protocol.

+-----------+-------+-----------+
| Command   | Bin   | Text      |
+-----------+-------+-----------+
| Get       | 0x30  | rget      |
| Set       | 0x31  | rset      |
| QSet      | 0x32  |           |
| Append    | 0x33  | rappend   |
| QAppend   | 0x34  |           |
| Prepend   | 0x35  | rprepend  |
| QPrepend  | 0x36  |           |
| Delete    | 0x37  | rdelete   |
| QDelete   | 0x38  |           |
| Incr      | 0x39  | rincr     |
| QIncr     | 0x3a  |           |
| Decr      | 0x3b  | rdecr     |
| QDecr     | 0x3c  |           |
+-----------+-------+-----------+

Isolation

Isolation levels are specifically left unspecified. Operations concurrently working on the same data set may interleave, mutually exclude each other, or just stomp all over each other.

Implementors are encouraged to document their intentions.

P.S.

Original: https://code.google.com/p/memcached/wiki/RangeOps Motive to duplicate - google source code (with project wikis) is closing.

Plan

lidaohang commented 1 year ago

Mark