Benchmarks
In yyjsonr: Fast 'JSON', 'NDJSON' and 'GeoJSON' Parser and Generator

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 6,
  fig.height = 4
)

library(yyjsonr)
library(bench)
library(ggplot2)
library(tidyr)
library(ggbeeswarm)
library(geojsonsf)
library(sf)

library(ndjson)

Benchmark overview

Benchmarking was done an Apple M2 Silicon.
Test-cases were drawn from other packages and other examples seen in the wild

From NDJSON file

ndjson_filename <- tempfile()
df <- head( nycflights13::flights, 1000)
jsonlite::stream_out(df, file(ndjson_filename), verbose = FALSE)

res06 <- bench::mark(
  ndjson   = ndjson::stream_in(ndjson_filename),
  jsonlite = jsonlite::stream_in(file(ndjson_filename), verbose = FALSE),
  jsonify  = jsonify::from_ndjson(ndjson_filename),
  # arrow    = arrow::read_json_arrow(ndjson_filename),
  yyjsonr  = yyjsonr::read_ndjson_file  (ndjson_filename),
  # relative = TRUE,
  check = FALSE
)

res06$benchmark <- 'From NDJSON File'
knitr::kable(res06[, 1:5])
plot(res06) + theme_bw() + theme(legend.position = 'none')

library(dplyr)
plot_df <- res06 %>%
  mutate(package = as.character(expression)) %>%
  rename(speed = `itr/sec`)

ggplot(plot_df) + 
  geom_col(aes(package, speed, fill = package), 
           position = position_dodge2(preserve = "single")) + 
  theme_bw(20) + 
  theme(legend.position = 'none') + 
  scale_fill_manual(values = c(rep(grey(0.5), 4), 'dodgerblue3')) +
  geom_hline(yintercept = 1, color = 'red', alpha = 0.5, linetype = 2) + 
  labs(
    x = NULL,
    y = "Factor speed increase over {jsonlite}",
    title = "Reading NDJSON.  Speed-up compared to {jsonlite}",
    subtitle = "Red line indicates reference implementation {jsonlite}"
  )

To NDJSON File

ndjson_filename <- tempfile()
df <- head( nycflights13::flights, 1000)

res07 <- bench::mark(
  jsonlite = jsonlite::stream_out(df, file(ndjson_filename), verbose = FALSE),
  yyjsonr  = yyjsonr::write_ndjson_file(df, ndjson_filename),
  check = FALSE
)

res07$benchmark <- 'To NDJSON File'
knitr::kable(res07[, 1:5])
plot(res07) + theme_bw() + theme(legend.position = 'none')

To NDJSON String

ndjson_filename <- tempfile()
df <- head( nycflights13::flights, 1000)

res08 <- bench::mark(
  jsonify  = jsonify::to_ndjson(df),
  jsonlite = jsonlite::stream_out(df, con = textConnection(NULL, "w"), name = "greg", local = TRUE, verbose = FALSE),
  yyjsonr  = yyjsonr::write_ndjson_str(df),
  check = FALSE
)

res08$benchmark <- 'To NDJSON String'
knitr::kable(res08[, 1:5])
plot(res08) + theme_bw() + theme(legend.position = 'none')

Summary

library(dplyr)
plot_df <- bind_rows(
  res06, res07, res08
)

plot_df$benchmark <- factor(
  plot_df$benchmark, 
  levels = unique(plot_df$benchmark)
)

plot_df <- plot_df %>% 
  mutate(
    package = as.character(expression),
    iters   = `itr/sec`,
    speed   = iters
  ) %>%
  select(benchmark, package, iters, speed)

plot_df <- plot_df %>%
  group_by(benchmark) %>%
  mutate(
    ref_speed = speed[which(package %in% c('jsonlite', 'geojsonsf'))],
    speed = speed / ref_speed
  ) %>%
  ungroup()

ggplot(plot_df) + 
  geom_col(aes(package, speed, fill = package), 
           position = position_dodge2(preserve = "single")) + 
  facet_wrap(~benchmark, scales = 'free_y', ncol = 3) + 
  theme_bw(15) + 
  theme(legend.position = 'none') + 
  scale_fill_manual(values = c(rep(grey(0.5), 3), 'dodgerblue3')) +
  geom_hline(yintercept = 1, color = 'red', alpha = 0.5, linetype = 2) + 
  labs(
    x = NULL,
    y = "Factor speed increase\nover reference implementation",
    title = "Speed-up compared to reference implementation",
    subtitle = "Red line indicates reference implementation {jsonlite}"
  ) +
  scale_y_continuous(breaks = scales::pretty_breaks())

if (FALSE) {
  ggsave("./man/figures/benchmark-ndjson.png", width = 9, height = 5)
  # saveRDS(plot_df, "man/benchmark/cache-df-types.rds")
}