---
title: "4. Splitting large queries"
author: 
  - "Mark Padgham"
  - "Martin Machyna"
date: "`r Sys.Date()`"
bibliography: osmdata-refs.bib
output: 
    html_document:
        toc: true
        toc_float: true
        number_sections: false
        theme: flatly
vignette: >
  %\VignetteIndexEntry{4. query-split}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

## 1. Introduction

The `osmdata` package retrieves data from the [`overpass`
server](https://overpass-api.de) which is primarily designed to deliver small
subsets of the full Open Street Map (OSM) data set, determined both by specific
bounding coordinates and specific OSM key-value pairs. The server has internal
routines to limit delivery rates on queries for excessively large data sets,
and may ultimately fail for large queries. This vignette describes one approach
for breaking overly large queries into a set of smaller queries, and for
re-combining the resulting data sets into a single `osmdata` object reflecting
the desired, large query.


## 2. Query splitting

Complex or data-heavy queries may exhaust the time or memory limits of the
`overpass` server. One way to get around this problem is to split the bounding
box (bbox) of a query into several smaller fragments, and then to re-combine
the data and remove duplicate objects. This section demonstrates how that may
be done, starting with a large bounding box.

```{r get-bbox, eval = FALSE}
library (osmdata)

bb <- getbb ("Southeastern Connecticut COG", featuretype = "boundary")
bb
```
```{r out1, eval = FALSE}
#>         min       max
#> x -72.46677 -71.79315
#> y  41.27591  41.75617
```

The following lines then divide that bounding box into two smaller areas:

```{r bbox-split, eval = FALSE}
dx <- (bb ["x", "max"] - bb ["x", "min"]) / 2

bbs <- list (bb, bb)

bbs [[1]] ["x", "max"] <- bb ["x", "max"] - dx
bbs [[2]] ["x", "min"] <- bb ["x", "min"] + dx

bbs
```
```{r out2, eval = FALSE}
#> [[1]]
#>         min       max
#> x -72.46677 -72.12996
#> y  41.27591  41.75617
#>
#> [[2]]
#>         min       max
#> x -72.12996 -71.79315
#> y  41.27591  41.75617
```

These two bounding boxes can then be used to submit two separate overpass
queries:

```{r opq-2x, eval = FALSE}
res <- list ()

res [[1]] <- opq (bbox = bbs [[1]]) |>
    add_osm_feature (key = "admin_level", value = "8") |>
    osmdata_sf ()
res [[2]] <- opq (bbox = bbs [[2]]) |>
    add_osm_feature (key = "admin_level", value = "8") |>
    osmdata_sf ()
```

The retrieved `osmdata` objects can then be merged using the`c(...)` function,
which automatically removes duplicate objects.

```{r opq-merge, eval = FALSE}
res <- c (res [[1]], res [[2]])
```


## 3. Automatic bbox splitting

The previous code demonstrated how to divide a bounding box into two, smaller
regions. It will generally not be possible to know in advance how small a
bounding box should be for a query for work, and so we need a more general
version of that functionality to divide a bounding box into a arbitrary number
of sub-regions.

We can automate this process by monitoring the exit status of `opq() |>
osmdata_sf()` and in case of a failed query we can keep recursively splitting
the current bounding box into increasingly smaller fragments until the overpass
server returns a result. The following function demonstrates splitting a
bounding box into a list of four equal-sized bounding boxes in a 2-by-2 grid,
each box having a specified degree of overlap (`eps=0.05`, or 5%) with the
neighbouring box.

```{r bbox-auto-split, eval = FALSE}
split_bbox <- function (bbox, grid = 2, eps = 0.05) {
    xmin <- bbox ["x", "min"]
    ymin <- bbox ["y", "min"]
    dx <- (bbox ["x", "max"] - bbox ["x", "min"]) / grid
    dy <- (bbox ["y", "max"] - bbox ["y", "min"]) / grid

    bboxl <- list ()

    for (i in 1:grid) {
        for (j in 1:grid) {
            b <- matrix (c (
                xmin + ((i - 1 - eps) * dx),
                ymin + ((j - 1 - eps) * dy),
                xmin + ((i + eps) * dx),
                ymin + ((j + eps) * dy)
            ),
            nrow = 2,
            dimnames = dimnames (bbox)
            )

            bboxl <- append (bboxl, list (b))
        }
    }
    bboxl
}
```

We pre-split our area and create a queue of bounding boxes that we will use for 
submitting queries.

```{r bbox-pre-split, eval = FALSE}
bb <- getbb ("Connecticut", featuretype = NULL)
queue <- split_bbox (bb)
result <- list ()
```

Now we can create a loop that will monitor the exit status of our query and in 
case of success remove the bounding box from the queue. If our query fails for 
some reason, we split the failed bounding box into four smaller fragments and
add them to our queue, repeating until all results have been successfully
delivered.

```{r auto-query, eval = FALSE}
while (length (queue) > 0) {

    print (queue [[1]])

    opres <- NULL
    opres <- try ({
        opq (bbox = queue [[1]], timeout = 25) |>
            add_osm_feature (key = "natural", value = "tree") |>
            osmdata_sf ()
    })

    if (class (opres) [1] != "try-error") {
        result <- append (result, list (opres))
        queue <- queue [-1]
    } else {
        bboxnew <- split_bbox (queue [[1]])
        queue <- append (bboxnew, queue [-1])
    }
}
```

All retrieved `osmdata` objects stored in the `result` list can then be
combined using the `c(...)` operator. Note that for large datasets this process
can be quite time consuming.

```{r merge-result-list, eval = FALSE}
final <- do.call (c, result)
```