README.md

In andrie/shimmer: Discrete Event Simulation for Scaling Shiny, Using the Simmer Package

shimmer

The shimmer package contains a discrete event simulation that explores how shiny processes behave at scale, typically orchestrated by RStudio Connect or Shiny Server Pro.

The underlying infrastructure of the simulation is provided by the simmer package (for discrete event simulations). In other words, shimmer simulates how Shiny apps scale by using the simmer simulation framework.

Installation

The package is not yet on CRAN…

… but you can install the development version from GitHub using:

# install.packages("devtools")
devtools::install_github("andrie/shimmer")

Example

The shimmer package uses discrete event simulation to help answer the questions:

• How big should my Shiny server be to handle n number of users?
• For a given size of server, how many users can Shiny handle?
• How should I tune the runtime parameters in RStudio Connect for my app?

The robust answer to this question is to use shinyloadtest, but for planning purposes you may want to rapidly develop some hypotheses and intuition about the problem, prior to building and testing an app.

Relationship between shimmer and the simmer package

The simmer package makes it easy to build discrete event simulations in R. The shimmer package uses simmer under the hood for defining and running the simulation.

Setting up the simulation

The shimmer() function reads a configuration file using the config::get() function. The package contains a default configuration file at:

system.file("config.yml", package = "shimmer")
#> [1] "C:/Users/apdev/Documents/R/win-library/3.4/shimmer/config.yml"

The contents of this file:

default:
  runtime:
    comment: The app runtime settings that are available in RStudio Connect
    max_processes: 3
    min_processes: 0
    max_connections_per_process: 20
    load_factor: 0.5
    idle_timeout_per_process: 5.0
    initial_timeout: 300
    connection_timeout: 3600
    read_timeout: 3600
  app:
    comment: Describes the app startup time and response time per click
    startup_time: 5.0
    response_time: 2.0
  user:
    arrival:
      comment: Arrival time between users (seconds)
      mean: 10.0
      shape: 5.0
    number_of_requests_per_user: 20.0
    request:
      comment: Mean arrival time between requests for a given user (seconds)
      mean: 10.0
      shape: 5.0
    idle:
      comment: Time in seconds that connection remains live after last request
      mean: 1800
      sd: 600
  system:
    cpu: 4.0

Run the simulation

library(magrittr)
library(shimmer)

By default, the simulation runs for an hour (3,600 seconds) of simulation time:

env <- shimmer()
#> You must specify either config or a valid config_file.
#> Using the built-in config file.
env
#> simmer environment: Shiny | now: 3600 | next: 3600
#> { Monitor: in memory }
#> { Resource: connection_request | monitored: TRUE | server status: 60(60) | queue status: 0(0) }
#> { Resource: total_connections | monitored: TRUE | server status: 123(Inf) | queue status: 0(0) }
#> { Resource: rejections | monitored: TRUE | server status: 243(Inf) | queue status: 0(0) }
#> { Resource: connection | monitored: TRUE | server status: 60(60) | queue status: 0(Inf) }
#> { Resource: cpu | monitored: TRUE | server status: 1(4) | queue status: 0(Inf) }
#> { Resource: process_1 | monitored: TRUE | server status: 20(20) | queue status: 0(0) }
#> { Resource: request_queue_1 | monitored: TRUE | server status: 0(1) | queue status: 0(Inf) }
#> { Resource: process_2 | monitored: TRUE | server status: 20(20) | queue status: 0(0) }
#> { Resource: request_queue_2 | monitored: TRUE | server status: 1(1) | queue status: 1(Inf) }
#> { Resource: process_3 | monitored: TRUE | server status: 20(20) | queue status: 0(0) }
#> { Resource: request_queue_3 | monitored: TRUE | server status: 0(1) | queue status: 0(Inf) }
#> { Source: controller | monitored: 1 | n_generated: 1 }
#> { Source: user_accounting | monitored: 1 | n_generated: 367 }

Plots

You can generate several plots from the simulation:

• CPU usage
• Response histogram
• Rejected connections (because the system was too busy)

env %>%
  plot_shimmer_cpu_usage()

env %>%
  plot_shimmer_response_histogram()

env %>%
  plot_shimmer_rejection_usage()

In addition, you can also get more detail of the underlying system behaviour:

• Connections
• Connections per process

env %>%
  plot_shimmer_connection_usage()

env %>%
  plot_shimmer_process_usage()

andrie/shimmer documentation built on May 3, 2019, 3:39 p.m.