saber233/redis-memory-analyzer

Redis memory profiler to find the RAM bottlenecks throw scaning key space in real time and aggregate RAM usage statistic by patterns.

Redis Memory Analyzer

RMA is a console tool to scan Redis key space in real time and aggregate memory usage statistic by key patterns. You may use this tools without maintenance on production servers. You can scanning by all or selected Redis types such as "string", "hash", "list", "set", "zset" and use matching pattern as you like. RMA try to discern key names by patterns, for example if you have keys like 'user:100' and 'user:101' application would pick out common pattern 'user:*' in output so you can analyze most memory distressed data in your instance.

Installing rma

Pre-Requisites :

1. python >= 3.5 and pip.
2. redis-py.

To install from PyPI (recommended) :

pip install rma

To install from source :

pip install git+https://github.com/gamenet/redis-memory-analyzer@v0.2.0

Running

After install used it from console:

>rma --help
usage: rma [-h] [-s HOST] [-p PORT] [-a PASSWORD] [-d DB] [-m MATCH] [-l LIMIT]
           [-b BEHAVIOUR] [-t TYPES]

RMA is used to scan Redis key space in and aggregate memory usage statistic by
key patterns.

optional arguments:
  -h, --help                 show this help message and exit
  -s, --server HOST          Redis Server hostname. Defaults to 127.0.0.1
  -p, --port PORT            Redis Server port. Defaults to 6379
  -a, --password PASSWORD    Password to use when connecting to the server
  -d, --db DB                Database number, defaults to 0
  -m, --match MATCH          Keys pattern to match
  -l, --limit LIMIT          Get max key matched by pattern
  -b, --behaviour BEHAVIOUR  Specify application working mode. Allowed values
                             are all, scanner, ram, global
  -t, --type TYPES           Data types to include. Possible values are string,
                             hash, list, set. Multiple types can be provided. If
                             not specified, all data types will be returned.
                             Allowed values arestring, hash, list, set, zset
  -f --format TYPE           Output type format: json or text (by default)
  -x --separator SEPARATOR   Specify namespace separator. Default is ':'

If you have large database try running first with --limit option to run first limited amount of keys. Also run with --types to limit only specified Redis types in large database. Not this tool has performance issues - call encoding for individual keys instead if batch queue with LUA (like in scanner does). So this option may be very useful. You can choose what kind of data would be aggregated from Redis node using -b (--behaviour) option as console argument. Supported behaviours are 'global', 'scanner', 'ram' and 'all'.

Internals

RMA shows statistics separated by types. All works in application separated by few steps:

1. Load type and encoding for each key matched by given pattern with Lua scripting in batch mode. SCAN used to iterate keys from Redis key db.
2. Separate keys by types and match patterns.
3. Run behaviours and rules for given data set.
4. Output result with given reported (now only TextReported implemented)

Global output ('global' behaviour)

The global data is some Redis server statistics which helps you to understand other data from this tools:

| Stat                             | Value          |
|:---------------------------------|:---------------|
| Total keys in db                 | 28979          |
| RedisDB key space overhead       | 790528         |
| Used `set-max-intset-entries`    | 512            |
| ....                             | ...            |
| Info `total_system_memory`       | 3190095872     |
| ....                             | ...            |

The one of interesting things here is "RedisDB key space overhead". The amount of memory used Redis to store key space data. If you have lots of keys in your Redis instance this actually shows your overhead for this. Keep in mind that part of data such as total keys in db or key space overhead shows data for selected db. But statistics started with Info or Config keywords is server based.

Key types ('scanner' behaviour)

This table helps then you do not know actually that kind of keys stored in your Redis database. For example then DevOps or system administrator want to understand what kind of keys stored in Redis instance. Which data structure is most used in system. This also helps if you are new to some big project - this kind of SHOW ALL TABLES request :)

| Match                 |   Count | Type   | %      |
|:----------------------|--------:|:-------|:-------|
| job:*                 |    5254 | hash   | 18.13% |
| game:privacy:*        |    2675 | hash   | 9.23%  |
| user:*                |    1890 | hash   | 6.52%  |
| group:*               |    1885 | set    | 6.50%  |

Data related output ('ram' behaviour)

All output separated by keys and values statistics. This division is used because: 1. Keys of any type in Redis actually stored in RedisDB internal data structure based on dict (more about this on RedisPlanet). 2. This type of data specially important in Redis instances with lots of keys.

| Match                         | Count | Useful |   Real | Ratio | Encoding                     | Min | Max |   Avg |
|:------------------------------|------:|-------:|-------:|------:|:-----------------------------|----:|----:|------:|
| event:data:*                  |  1198 |  17970 |  76672 |  4.27 | embstr [50.0%] / raw [50.0%] |  15 |  71 | 41.20 |
| mm:urllist:*                  |   524 |   7648 |  33536 |  4.38 | embstr [100.0%]              |  12 |  15 | 14.60 |
| Provider:ParallelForm:*:*:*:* |   459 |  43051 |  66096 |  1.54 | raw [100.0%]                 |  92 |  94 | 93.79 |
| user:spamblocked:dialy:post:* |    48 |   2208 |   4608 |  2.09 | raw [100.0%]                 |  46 |  46 | 46.00 |
| ...                           |   ... |    ... |    ... |   ... |                          ... | ... | ... |   ... |
| Total:                        |  2432 |  80493 | 200528 |  0.00 |                              |   0 |   0 |  0.00 |

So you can see count of keys matching given pattern, expected (by developer) and real memory with taking into account the Redis data structures and allocator overhead. Ratio and encoding distribution min/max/avg len of key. For example in sample above keys some keys encoded as raw (sds string). Each sds encoded string:

1. Has useful payload
2. Has sds string header overhead
3. Has redis object overhead
4. The Redis implementation during memory allocation would be align(redis object) + align(sds header + useful payload)

In x64 instance of Redis key event:data:f1wFFqgqqwgeg (24 byte len) actually would use 24 bytes payload bytes, 9 bytes sds header and 32 bytes in r_obj (redis object). So we may think this would use 65 bytes. But after jemalloc allocator align it this 24 byte (65 byte data with Redis internals) would use 80 bytes - in ~3,3 more times as you expect (`Ratio`` value in table).

Not we can look at values. All values output individual by Redis type. Each type has they own limitations so here is some common data for each type and some unique. The strings data type value same as keys output above. The only one difference is Free field which shows unused but allocated memory by SDS strings in raw encoding.

So for example look at output for HASH values:

| Match                 | Count | Avg field count | Key mem |   Real | Ratio | Value mem |   Real |    Ratio |   System | Encoding         | Total mem |  Total aligned |
|:----------------------|------:|----------------:|--------:|-------:|------:|----------:|-------:|---------:|---------:|:-----------------|----------:|---------------:|
| job:*                 |  5254 |            9.00 |  299485 | 619988 |  2.07 |    685451 | 942984 |     1.38 |  1345024 | ziplist [100.0%] |    984936 |        2907996 |
| LIKE:*                |  1890 |            1.02 |    5744 |  30262 |  5.27 |      1932 |  15432 |     7.99 |    91344 | ziplist [100.0%] |      7676 |         137038 |
| game:*:count:*        |  1231 |            1.00 |    7386 |  19696 |  2.67 |      1234 |   9848 |     7.98 |    59088 | ziplist [100.0%] |      8620 |          88632 |
| LIKE:game:like:*      |  1207 |            1.00 |    3621 |  19312 |  5.33 |      1210 |   9656 |     7.98 |    57936 | ziplist [100.0%] |      4831 |          86904 |
| integration:privacy:* |   530 |            3.00 |   20140 |  33920 |  1.68 |         0 |  25440 | 25440.00 |    42400 | ziplist [100.0%] |     20140 |         101760 |

Look at job:* hashes. This instance contains 5254 such keys with 9 fields each. Looks like this data has regular structure like python tuple. This means you can change data structure of this data from Redis hash to list and use 2 times less memory then now. Why do this? Now you job:* hash uses ~3,2 times more memory as you developers expect.

Why doesn't reported memory match actual memory used?

The memory reported by this tool is approximate. In general, the reported memory should be within 10% of what is reported by info.

Also note that the tool does not (and cannot) account for the following: - Memory used by allocator metadata (it is actually not possible without c) - Memory used for pub/sub (no any commands in Redis for that) - Redis process internals (like shared objects)

Known issues

1. Skiplist (zset actually) encoding actually not realized.
2. Quicklist now calculated as ziplist.
3. SDS strings from redis 3.2 (optimized headers) not implemented. Now used fixed 9 bytes header.

Whats next?

Now we use this tools as awesome helper. We most used data structures in our Redis instances is hash and list. After upgradings our servers to Redis 3.2.x planning to fix known issues. Be glad to know that are you think about this tool. In my dreams this tools should used as redis-lint tools which can say you Hey, change this from this to this and save 30% of RAM, Hey, you are using PHP serializer for strings - change to msgpack and save 15% of RAM and so on.

License

This application was developed for using in GameNet project as part of Redis memory optimizations and analise. RMA is licensed under the MIT License. See LICENSE