Learn R Programming

gargle (version 1.2.1)

request_retry: Make a Google API request, repeatedly

Description

Intended primarily for internal use in client packages that provide high-level wrappers for users. It is a drop-in substitute for request_make() that also has the ability to retry the request. Codes that are considered retryable: 408, 429, 500, 502, 503.

Usage

request_retry(..., max_tries_total = 5, max_total_wait_time_in_seconds = 100)

Value

Object of class response from httr.

Arguments

...

Passed along to request_make().

max_tries_total

Maximum number of tries.

max_total_wait_time_in_seconds

Total seconds we are willing to dedicate to waiting, summed across all tries. This is a technical upper bound and actual cumulative waiting will be less.

Special cases

request_retry() departs from exponential backoff in three special cases:

  • It actually implements truncated exponential backoff. There is a floor and a ceiling on random wait times.

  • Retry-After header: If the response has a header named Retry-After (case-insensitive), it is assumed to provide a non-negative integer indicating the number of seconds to wait. If present, we wait this many seconds and do not generate a random waiting time. (In theory, this header can alternatively provide a datetime after which to retry, but we have no first-hand experience with this variant for a Google API.)

  • Sheets API quota exhaustion: In the course of googlesheets4 development, we've grown very familiar with the 429 RESOURCE_EXHAUSTED error. The Sheets API v4 has "a limit of 500 requests per 100 seconds per project and 100 requests per 100 seconds per user. Limits for reads and writes are tracked separately." In our experience, the "100 (read or write) requests per 100 seconds per user" limit is the one you hit most often. If we detect this specific failure, the first wait time is a bit more than 100 seconds, then we revert to exponential backoff.

Details

Consider an example where we are willing to make a request up to 5 times.

try  1  2    3        4                5
     |--|----|--------|----------------|
wait  1   2      3           4

There will be up to 5 - 1 = 4 waits and we generally want the waiting period to get longer, in an exponential way. Such schemes are called exponential backoff. request_retry() implements exponential backoff with "full jitter", where each waiting time is generated from a uniform distribution, where the interval of support grows exponentially. A common alternative is "equal jitter", which adds some noise to fixed, exponentially increasing waiting times.

Either way our waiting times are based on a geometric series, which, by convention, is usually written in terms of powers of 2:

b , 2b, 4b, 8b, ...
  = b * 2^0, b * 2^1, b * 2^2, b * 2^3, ...

The terms in this series require knowledge of b, the so-called exponential base, and many retry functions and libraries require the user to specify this. But most users find it easier to declare the total amount of waiting time they can tolerate for one request. Therefore request_retry() asks for that instead and solves for b internally. This is inspired by the Opnieuw Python library for retries. Opnieuw's interface is designed to eliminate uncertainty around:

  • Units: Is this thing given in seconds? minutes? milliseconds?

  • Ambiguity around how things are counted: Are we starting at 0 or 1? Are we counting tries or just the retries?

  • Non-intuitive required inputs, e.g., the exponential base.

Let n be the total number of tries we're willing to make (the argument max_tries_total) and let W be the total amount of seconds we're willing to dedicate to making and retrying this request (the argument max_total_wait_time_in_seconds). Here's how we determine b:

sum_{i=0}^(n - 1) b * 2^i = W
b * sum_{i=0}^(n - 1) 2^i = W
       b * ( (2 ^ n) - 1) = W
                        b = W / ( (2 ^ n) - 1)

See Also

Examples

Run this code
if (FALSE) {
req <- gargle::request_build(
  method = "GET",
  path = "path/to/the/resource",
  token = "PRETEND_I_AM_TOKEN"
)
gargle::request_retry(req)
}

Run the code above in your browser using DataLab