Cache R Objects with Automatic Pruning.

• cachem
  • Installation
  • Usage
  • Cache types
    • cache_mem()
    • cache_disk()
  • Cache API
  • Pruning
  • Layered caches

cachem

The cachem R package provides objects creating and managing caches. These cache objects are key-value stores, but unlike other basic key-value stores, they have built-in support for memory and age limits so that they won’t have unbounded growth.

The cache objects in cachem differ from some other key-value stores in the following ways:

• The cache objects provide automatic pruning so that they remain within memory limits.
• Fetching a non-existing object returns a sentinel value. An alternative is to simply return NULL. This is what R lists and environments do, but it is ambiguous whether the value really is NULL, or if it is not present. Another alternative is to throw an exception when fetching a non-existent object. However, this results in more complicated code, as every get() needs to be wrapped in a tryCatch().

Installation

To install the CRAN version:

install.packages("cachem")

You can install the development version from with:

if (!require("remotes")) install.packages("remotes")
remotes::install_github("r-lib/cachem")

Usage

To create a memory-based cache, call cache_mem().

library(cachem)
m <- cache_mem()

Add arbitrary R objects to the cache using $set(key, value):

m$set("abc123", c("Hello", "world"))
m$set("xyz", function() message("Goodbye"))

The key must be a string consisting of lowercase letters, numbers, and the underscore (_) and hyphen (-) characters. (Upper-case characters are not allowed because some storage backends do not distinguish between lowercase and uppercase letters.) The value can be any R object.

Get the values with $get():

m$get("abc123")
#> [1] "Hello" "world"

m$get("xyz")
#> function() message("Goodbye")

If you call get() on a key that doesn’t exists, it will return a key_missing() sentinel value:

m$get("dog")
#> <Key Missing>

A common usage pattern is to call get(), and then check if the result is a key_missing object:

value <- m$get(key)

if (is.key_missing(value)) {
  # Cache miss - do something
} else {
  # Cache hit - do another thing
}

The reason for doing this (instead of calling $exists(key) and then $get(key)) is that for some storage backends, there is a potential race condition: the object could be removed from the cache between the exists() and get() calls. For example:

• If multiple R processes have cache_disks that share the same directory, one process could remove an object from the cache in between the exists() and get() calls in another process, resulting in an error.
• If you use a cache_mem with a max_age, it’s possible for an object to be present when you call exists(), but for its age to exceed max_age by the time get() is called. In that case, the get() will return a key_missing() object.

# Avoid this pattern, due to a potential race condition!
if (m$exists(key)) {
  value <- m$get(key)
}

Cache types

cachem comes with two kinds of cache objects: a memory cache, and a disk cache.

cache_mem()

The memory cache stores stores objects in memory, by simply keeping a reference to each object. To create a memory cache:

m <- cache_mem()

The default size of the cache is 200MB, but this can be customized with max_size:

m <- cache_mem(max_size = 10 * 1024^2)

It may also be useful to set a maximum age of objects. For example, if you only want objects to stay for a maximum of one hour:

m <- cache_mem(max_size = 10 * 1024^2, max_age = 3600)

For more about how objects are evicted from the cache, see section Pruning below.

An advantage that the memory cache has over the disk cache (and any other type of cache that stores the objects outside of the R process’s memory), is that it does not need to serialize objects. Instead, it merely stores references to the objects. This means that it can store objects that other caches cannot, and with more efficient use of memory – if two objects in the cache share some of their contents (such that they refer to the same sub-object in memory), then cache_mem will not create duplicate copies of the contents, as cache_disk would, since it serializes the objects with the serialize() function.

Compared to the memory usage, the size calculation is not as intelligent: if there are two objects that share contents, their sizes are computed separately, even if they have items that share the exact same represention in memory. This is done with the object.size() function, which does not account for multiple references to the same object in memory.

In short, a memory cache, if anything, over-counts the amount of memory actually consumed. In practice, this means that if you set a 200MB limit to the size of cache, and the cache thinks it has 200MB of contents, the actual amount of memory consumed could be less than 200MB.

# Create a and b which both contain the same numeric vector.
x <- list(rnorm(1e5))
a <- list(1, x)
b <- list(2, x)

# Add to cache
m$set("a", a)
m$set("b", b)

# Each object is about 800kB in memory, so the cache_mem() will consider the
# total memory used to be 1600kB.
object.size(m$get("a"))
#> 800224 bytes
object.size(m$get("b"))
#> 800224 bytes

lobstr::obj_size(m$get("a"))
#> 800.22 kB
lobstr::obj_size(list(m$get("a"), m$get("b")))
#> 800.41 kB

cache_disk()

Disk caches are stored in a directory on disk. A disk cache is slower than a memory cache, but can generally be larger. To create one:

d <- cache_disk()

By default, it creates a subdirectory of the R process’s temp directory, and it will persist until the R process exits.

d$info()$dir
#>  "/tmp/Rtmp6h5iB3/cache_disk-d1901b2b615a"

Like a cache_mem, the max_size, max_n, max_age can be customized. See section Pruning below for more information.

Each object in the cache is stored as an RDS file on disk, using the serialize() function.

d$set("abc", 100)
d$set("x01", list(1, 2, 3))

dir(d$info()$dir)
#> [1] "abc.rds" "x01.rds"

Since objects in a disk cache are serialized, they are subject to the limitations of the serialize() function. For more information, see section Limitations of serialized objects.

The storage directory can be specified with dir; it will be created if necessary.

cache_disk(dir = "cachedir")

Sharing a disk cache among processes

Multiple R processes can use disk_cache objects that share the same cache directory. To do this, simply point each cache_disk to the same directory.

disk_cache pruning

For a disk_cache, pruning does not happen on every access, because finding the size of files in the cache directory can take a nontrivial amount of time. By default, pruning happens once every 20 times that $set() is called, or if at least five seconds have elapsed since the last pruning. The prune_rate controls how many times $set() must be called before a pruning occurs. It defaults to 20; smaller values result in more frequent pruning and larger values result in less frequent pruning (but keep in mind pruning always occurs if it has been at least five seconds since the last pruning).

Cleaning up the cache directory

The cache directory can be deleted by calling $destroy(). After it is destroyed, the cache object can no longer be used.

d$destroy()
d$set("a", 1)  # Error

To create a cache_disk that will automatically delete its storage directory when garbage collected, use destroy_on_finalize=TRUE:

d <- cache_disk(destroy_on_finalize = TRUE)
d$set("a", 1)

cachedir <- d$info()$dir
dir(cachedir)
#> [1] "a.rds"

# Remove reference to d and trigger a garbage collection
rm(d)
gc()

dir.exists(cachedir)

Using custom serialization functions

It is possible to use custom serialization functions rather than the default of writeRDS() and readRDS() with the write_fn, read_fn and extension arguments respectively. This could be used to use alternative serialization formats like qs, or specialized object formats fst or parquet.

library(qs)

d <- cache_disk(read_fn = qs::qread, write_fn = qs::qsave, extension = ".qs")

d$set("a", list(1, 2, 3))

cachedir <- d$info()$dir
dir(cachedir)
#> [1] "a.qs"
d$get("a")
#> [[1]]
#> [1] 1
#>
#> [[2]]
#> [1] 2
#>
#> [[3]]
#> [1] 3

Cache API

cache_mem() and cache_disk() support all of the methods listed below. If you want to create a compatible caching object, it must have at least the get() and set() methods:

• get(key, missing = missing_): Get the object associated with key. The missing parameter allows customized behavior if the key is not present: it actually is an expression which is evaluated when there is a cache miss, and it could return a value or throw an error.
• set(key, value): Set a key to a value.
• exists(key): Check whether a particular key exists in the cache.
• remove(key): Remove a key-value from the cache.

Some optional methods:

• reset(): Clear all objects from the cache.
• keys(): Return a character vector of all keys in the cache.
• prune(): Prune the cache. (Some types of caches may not prune on every access, and may temporarily grow past their limits, until the next pruning is triggered automatically, or manually with this function.)
• size(): Return the number of objects in the cache.

For these methods:

• key: can be any string with lowercase letters, numbers, underscore (_) and hyphen (-). Some storage backends may not be handle very long keys well. For example, with a cache_disk(), the key is used as a filename, and on some filesystems, very filenames may hit limits on path lengths.
• value: can be any R object, with some exceptions noted below.

Limitations of serialized objects

For any cache that serializes the object for storage outside of the R process – in other words, any cache other than a cache_mem() – some types of objects will not save and restore as well. Notably, reference objects may consume more memory when restored, since R may not know to deduplicate shared objects. External pointers are not be able to be serialized, since they point to memory in the R process. See ?serialize for more information.

Read-only caches

It is possible to create a read-only cache by making the set(), remove(), reset(), and prune() methods into no-ops. This can be useful if sharing a cache with another R process which can write to the cache. For example, one (or more) processes can write to the cache, and other processes can read from it.

This function will wrap a cache object in a read-only wrapper. Note, however, that code that uses such a cache must not require that $set() actually sets a value in the cache. This is good practice anyway, because with these cache objects, items can be pruned from them at any time.

cache_readonly_wrap <- function(cache) {
  structure(
    list(
      get = cache$get,
      set = function(key, value) NULL,
      exists = cache$exists,
      keys = cache$keys,
      remove = function(key) NULL,
      reset = function() NULL,
      prune = function() NULL,
      size = cache$size
    ),
    class = c("cache_readonly", class(cache))
  )
}

mr <- cache_readonly_wrap(m)

Pruning

The cache objects provided by cachem have automatic pruning. (Note that pruning is not required by the API, so one could implement an API-compatible cache without pruning.)

This section describes how pruning works for cache_mem() and cache_disk().

When the cache object is created, the maximum size (in bytes) is specified by max_size. When the size of objects in the cache exceeds max_size, objects will be pruned from the cache.

When objects are pruned from the cache, which ones are removed is determined by the eviction policy, evict:

• lru: The least-recently-used objects will be removed from the cache, until it fits within the limit. This is the default and is appropriate for most cases.
• fifo: The oldest objects will be removed first.

It is also possible to set the maximum number of items that can be in the cache, with max_n. By default this is set to Inf, or no limit.

The max_age parameter is somewhat different from max_size and max_n. The latter two set limits on the cache store as a whole, whereas max_age sets limits for each individual item; for each item, if its age exceeds max_age, then it will be removed from the cache.

Layered caches

Multiple caches can be composed into a single cache, using cache_layered(). This can be used to create a multi-level cache. (Note thate cache_layered() is currently experimental.) For example, we can create a layered cache with a very fast 100MB memory cache and a larger but slower 2GB disk cache:

m <- cache_mem(max_size = 100 * 1024^2)
d <- cache_disk(max_size = 2 * 1024^3)

cl <- cache_layered(m, d)

The layered cache will have the same API, with $get(), $set(), and so on, so it can be used interchangeably with other caching objects.

For this example, we’ll recreate the cache_layered with logging enabled, so that it will show cache hits and misses.

cl <- cache_layered(m, d, logfile = stderr())

# Each of the objects generated by rnorm() is about 40 MB
cl$set("a", rnorm(5e6))
cl$set("b", rnorm(5e6))
cl$set("c", rnorm(5e6))

# View the objects in each of the component caches
m$keys()
#> [1] "c" "b"
d$keys()
#> [1] "a" "b" "c"

# The layered cache reports having all keys
cl$keys()
#> [1] "c" "b" "a"

When $get() is called, it searches the first cache, and if it’s missing there, it searches the next cache, and so on. If not found in any caches, it returns key_missing().

# Get object that exists in the memory cache
x <- cl$get("c")
#> [2020-10-23 13:11:09.985] cache_layered Get: c
#> [2020-10-23 13:11:09.985] cache_layered Get from cache_mem... hit

# Get object that doesn't exist in the memory cache
x <- cl$get("a")
#> [2020-10-23 13:13:10.968] cache_layered Get: a
#> [2020-10-23 13:13:10.969] cache_layered Get from cache_mem... miss
#> [2020-10-23 13:13:11.329] cache_layered Get from cache_disk... hit

# Object is not present in any component caches
cl$get("d")
#> [2020-10-23 13:13:40.197] cache_layered Get: d
#> [2020-10-23 13:13:40.197] cache_layered Get from cache_mem... miss
#> [2020-10-23 13:13:40.198] cache_layered Get from cache_disk... miss
#> <Key Missing>

Multiple cache objects can be layered this way. You could even add a cache which uses a remote store, such as a network file system or even AWS S3.