brycewang-stanford/r-performance

R Performance Best Practices

Profiling, benchmarking, and optimization strategies for R code

Performance Tool Selection Guide

When to Use Each Performance Tool

Profiling Tools Decision Matrix

| Tool | Use When | Don't Use When | What It Shows | |------|----------|----------------|---------------| | profvis | Complex code, unknown bottlenecks | Simple functions, known issues | Time per line, call stack | | bench::mark() | Comparing alternatives | Single approach | Relative performance, memory | | system.time() | Quick checks | Detailed analysis | Total runtime only | | Rprof() | Base R only environments | When profvis available | Raw profiling data |

Step-by-Step Performance Workflow

# 1. Profile first - find the actual bottlenecks
library(profvis)
profvis({
  # Your slow code here
})

# 2. Focus on the slowest parts (80/20 rule)
# Don't optimize until you know where time is spent

# 3. Benchmark alternatives for hot spots
library(bench)
bench::mark(
  current = current_approach(data),
  vectorized = vectorized_approach(data),
  parallel = map(data, in_parallel(func))
)

# 4. Consider tool trade-offs based on bottleneck type

When Each Tool Helps vs Hurts

Parallel Processing (in_parallel())

# Helps when:
# - CPU-intensive computations
# - Embarassingly parallel problems
# - Large datasets with independent operations
# - I/O bound operations (file reading, API calls)

# Hurts when:
# - Simple, fast operations (overhead > benefit)
# - Memory-intensive operations (may cause thrashing)
# - Operations requiring shared state
# - Small datasets

# Example decision point:
expensive_func <- function(x) Sys.sleep(0.1) # 100ms per call
fast_func <- function(x) x^2                 # microseconds per call

# Good for parallel
map(1:100, in_parallel(expensive_func))  # ~10s -> ~2.5s on 4 cores

# Bad for parallel (overhead > benefit)
map(1:100, in_parallel(fast_func))       # 100us -> 50ms (500x slower!)

vctrs Backend Tools

# Use vctrs when:
# - Type safety matters more than raw speed
# - Building reusable package functions
# - Complex coercion/combination logic
# - Consistent behavior across edge cases

# Avoid vctrs when:
# - One-off scripts where speed matters most
# - Simple operations where base R is sufficient
# - Memory is extremely constrained

# Decision point:
simple_combine <- function(x, y) c(x, y)           # Fast, simple
robust_combine <- function(x, y) vec_c(x, y)      # Safer, slight overhead

# Use simple for hot loops, robust for package APIs

Data Backend Selection

# Use data.table when:
# - Very large datasets (>1GB)
# - Complex grouping operations
# - Reference semantics desired
# - Maximum performance critical

# Use dplyr when:
# - Readability and maintainability priority
# - Complex joins and window functions
# - Team familiarity with tidyverse
# - Moderate sized data (<100MB)

# Use base R when:
# - No dependencies allowed
# - Simple operations
# - Teaching/learning contexts

Profiling Best Practices

# 1. Profile realistic data sizes
profvis({
  # Use actual data size, not toy examples
  real_data |> your_analysis()
})

# 2. Profile multiple runs for stability
bench::mark(
  your_function(data),
  min_iterations = 10,  # Multiple runs
  max_iterations = 100
)

# 3. Check memory usage too
bench::mark(
  approach1 = method1(data),
  approach2 = method2(data),
  check = FALSE,  # If outputs differ slightly
  filter_gc = FALSE  # Include GC time
)

# 4. Profile with realistic usage patterns
# Not just isolated function calls

Performance Anti-Patterns to Avoid

# Don't optimize without measuring
# BAD: "This looks slow" -> immediately rewrite
# GOOD: Profile first, optimize bottlenecks

# Don't over-engineer for performance
# BAD: Complex optimizations for 1% gains
# GOOD: Focus on algorithmic improvements

# Don't assume - measure
# BAD: "for loops are always slow in R"
# GOOD: Benchmark your specific use case

# Don't ignore readability costs
# BAD: Unreadable code for minor speedups
# GOOD: Readable code with targeted optimizations

Backend Tools for Performance

• Consider lower-level tools when speed is critical
• Use vctrs, rlang backends when appropriate
• Profile to identify true bottlenecks

# For packages - consider backend tools
# vctrs for type-stable vector operations
# rlang for metaprogramming
# data.table for large data operations

When to Use vctrs

Core Benefits

• Type stability - Predictable output types regardless of input values
• Size stability - Predictable output sizes from input sizes
• Consistent coercion rules - Single set of rules applied everywhere
• Robust class design - Proper S3 vector infrastructure

Use vctrs when

Building Custom Vector Classes

# Good - vctrs-based vector class
new_percent <- function(x = double()) {
  vec_assert(x, double())
  new_vctr(x, class = "pkg_percent")
}

# Automatic data frame compatibility, subsetting, etc.

Type-Stable Functions in Packages

# Good - Guaranteed output type
my_function <- function(x, y) {
  # Always returns double, regardless of input values
  vec_cast(result, double())
}

# Avoid - Type depends on data
sapply(x, function(i) if(condition) 1L else 1.0)

Consistent Coercion/Casting

# Good - Explicit casting with clear rules
vec_cast(x, double())  # Clear intent, predictable behavior

# Good - Common type finding
vec_ptype_common(x, y, z)  # Finds richest compatible type

# Avoid - Base R inconsistencies
c(factor("a"), "b")  # Unpredictable behavior

Size/Length Stability

# Good - Predictable sizing
vec_c(x, y)  # size = vec_size(x) + vec_size(y)
vec_rbind(df1, df2)  # size = sum of input sizes

# Avoid - Unpredictable sizing
c(env_object, function_object)  # Unpredictable length

vctrs vs Base R Decision Matrix

| Use Case | Base R | vctrs | When to Choose vctrs | |----------|--------|-------|---------------------| | Simple combining | c() | vec_c() | Need type stability, consistent rules | | Custom classes | S3 manually | new_vctr() | Want data frame compatibility, subsetting | | Type conversion | as.*() | vec_cast() | Need explicit, safe casting | | Finding common type | Not available | vec_ptype_common() | Combining heterogeneous inputs | | Size operations | length() | vec_size() | Working with non-vector objects |

Implementation Patterns

Basic Vector Class

# Constructor (low-level)
new_percent <- function(x = double()) {
  vec_assert(x, double())
  new_vctr(x, class = "pkg_percent")
}

# Helper (user-facing)
percent <- function(x = double()) {
  x <- vec_cast(x, double())
  new_percent(x)
}

# Format method
format.pkg_percent <- function(x, ...) {
  paste0(vec_data(x) * 100, "%")
}

Coercion Methods

# Self-coercion
vec_ptype2.pkg_percent.pkg_percent <- function(x, y, ...) {
  new_percent()
}

# With double
vec_ptype2.pkg_percent.double <- function(x, y, ...) double()
vec_ptype2.double.pkg_percent <- function(x, y, ...) double()

# Casting
vec_cast.pkg_percent.double <- function(x, to, ...) {
  new_percent(x)
}
vec_cast.double.pkg_percent <- function(x, to, ...) {
  vec_data(x)
}

Performance Considerations

When vctrs Adds Overhead

• Simple operations - vec_c(1, 2) vs c(1, 2) for basic atomic vectors
• One-off scripts - Type safety less critical than speed
• Small vectors - Overhead may outweigh benefits

When vctrs Improves Performance

• Package functions - Type stability prevents expensive re-computation
• Complex classes - Consistent behavior reduces debugging
• Data frame operations - Robust column type handling
• Repeated operations - Predictable types enable optimization

Package Development Guidelines

Exports and Dependencies

# DESCRIPTION - Import specific functions
Imports: vctrs

# NAMESPACE - Import what you need
importFrom(vctrs, vec_assert, new_vctr, vec_cast, vec_ptype_common)

# Or if using extensively
import(vctrs)

Testing vctrs Classes

# Test type stability
test_that("my_function is type stable", {
  expect_equal(vec_ptype(my_function(1:3)), vec_ptype(double()))
  expect_equal(vec_ptype(my_function(integer())), vec_ptype(double()))
})

# Test coercion
test_that("coercion works", {
  expect_equal(vec_ptype_common(new_percent(), 1.0), double())
  expect_error(vec_ptype_common(new_percent(), "a"))
})

Don't Use vctrs When

• Simple one-off analyses - Base R is sufficient
• No custom classes needed - Standard types work fine
• Performance critical + simple operations - Base R may be faster
• External API constraints - Must return base R types

The key insight: vctrs is most valuable in package development where type safety, consistency, and extensibility matter more than raw speed for simple operations.

Performance Migrations

# Old -> New performance patterns
for loops for parallelizable work -> map(data, in_parallel(f))
Manual type checking             -> vec_assert() / vec_cast()
Inconsistent coercion           -> vec_ptype_common() / vec_c()