For some functionality one single core plumber instance might not be
enough to achieve the performance that is desired. One way around this
is to use some cluster orchestration tools. The example below discusses
the usage of docker-compose
but other tools like
docker swarm
or kubernetes
should be able to
achieve similar results. Alternatively local parallelization can be
used.
We first start two tile servers running on port 4001
and
4002
.
require(callr)
rp_list <- lapply(lapply(as.list(4000+1:2), c, list(tmpGridFile=tmpGridFile)), r_bg, func=function(port, tmpGridFile) {
# read a stars grid
weatherData <- stars::read_stars(tmpGridFile, proxy = FALSE, sub = "t")
names(weatherData) <- "t"
sf::st_crs(weatherData) <- "+proj=longlat"
colorFunction <- leaflet::colorNumeric("viridis", c(250, 310))
colorFunctionWithAlpa <- function(x, alpha = 1) {
paste0(colorFunction(x), as.character(as.raw(
as.numeric(alpha) * 255
)))
}
starsTileServer::starsTileServer$new(weatherData, colorFun = colorFunctionWithAlpa)$run(port = port)
})
Now we can use the subdomains argument of addTiles
to
address both servers.
require(leaflet)
require(leaflet.extras)
map <- leaflet() %>%
addTiles() %>%
enableTileCaching() %>%
addTiles(
"http://127.0.0.1:400{s}/map/t/{z}/{x}/{y}?level=900&time=2000-04-27 01:00:00&alpha=0.5",
options = tileOptions(useCache = TRUE, crossOrigin = TRUE, subdomains = '12')
) %>%
setView(zoom = 3, lat = 30, lng = 30)
This map looks as follows:
Using lapply
we can close both servers.
An alternative approach is to use docker (or some similar functionality). This allows you to scale much broader and is probably an approach that is more suitable for large scale permanent deployments.
The first step is to build a docker image that can be used to set up
the service. This docker image runs the tileserver. A simple example of
a possible Dockerfile
could look as follows.
FROM rocker/geospatial
MAINTAINER Bart
RUN install2.r -n 5 plumber stars; \
rm -rf /tmp/downloaded_packages
RUN R --quiet -e 'install.packages("starsdata", repos = "http://pebesma.staff.ifgi.de", type = "source")'
RUN R --quiet -e "remotes::install_gitlab('bartk/starsTileServer')"
EXPOSE 3436
COPY script.R script.R
RUN R --quiet -e "source('script.R')"
ENTRYPOINT ["R", "--quiet", "-e", "server<-readRDS('server.rds') ;server$run( port=3436, host='0.0.0.0', swagger=T)"]
The following R script is used (script.R
):
require(stars)
require(starsTileServer)
s5p <- system.file(
"sentinel5p/S5P_NRTI_L2__NO2____20180717T120113_20180717T120613_03932_01_010002_20180717T125231.nc",
package = "starsdata"
)
nit <- read_stars(
s5p,
along = NA,
sub = c(
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/nitrogendioxide_total_column",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/nitrogendioxide_total_column_precision",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/nitrogendioxide_total_column_precision_kernel",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/number_of_iterations",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/number_of_spectral_points_in_retrieval",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/oxygen_oxygen_dimer_slant_column_density",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/oxygen_oxygen_dimer_slant_column_density_precision",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/ozone_slant_column_density",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/ozone_slant_column_density_precision",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/processing_quality_flags",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/ring_coefficient",
"//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/ring_coefficient_precision"
),
curvilinear = c("//PRODUCT/longitude", "//PRODUCT/latitude"),
driver = NULL
)
names(nit) <-
sub("//PRODUCT/SUPPORT_DATA/DETAILED_RESULTS/", "", names(nit))
for (i in seq(length(names(nit)))) {
nit[[i]][nit[[i]] > 9e+36] <- NA
}
st_crs(nit) <- 4326
server <- starsTileServer$new(nit)
# we save the server here as there should only be one version (sampling of color scales would otherwise result in differently colored tiles)
saveRDS(server, "server.rds")
Copies of these files can be found with the following commands:
With the following docker-compose.yml
file we can then
start the applications:
version: "2.2"
services:
tileserver:
build:
dockerfile: Dockerfile
context: .
scale: 4
restart: always
lb:
container_name: haproxy_tile_loadbalancing
image: 'dockercloud/haproxy:latest'
environment:
- TIMEOUT=connect 4000, client 153000, server 230000
links:
- tileserver
volumes:
- /var/run/docker.sock:/var/run/docker.sock
varnish:
image: wodby/varnish
container_name: varnish_tile_caching
ports:
- "80:80"
- "6081:6081"
- "8080:8080"
depends_on:
- lb
environment:
VARNISH_IMPORT_MODULES: cookie,header
VARNISH_CONFIG_PRESET: drupal
VARNISH_BACKEND_HOST: lb
VARNISH_BACKEND_PORT: 80
In this case 4 parallel instances are started. We use
haproxy
to distribute the load across the containers and
varnish
to cache the tiles that have been rendered before.
The caching makes sure no double work is done.
With the docker-compose build
command the required
docker containers can be build. Using docker-compose
up the
cluster can then be started. Now in normal R we can plot a leaflet maps
as was done before.