pkgdown/mathjax-config.html

Skip to contents

Introduction

Most analytical workflows combine data from multiple sources. Observations arrive in one file, site metadata lives in another, and a taxonomic reference list sits in a third. Joins are the operation that stitches them together.

vectra implements joins as C-level hash join nodes. Like every other verb, a join call builds a plan node; nothing executes until collect() pulls batches through the tree. The engine follows a build-right, probe-left model. When execution starts, the right-side table is fully materialized into a hash table in memory. The left side then streams through batch by batch, probing the hash table for matches. This asymmetry has a practical consequence: the left side can be arbitrarily large (it never lives in memory all at once), but the right side must fit. So we put the smaller table on the right.

vectra provides seven join verbs: left_join(), inner_join(), right_join(), full_join(), semi_join(), anti_join(), and cross_join(). For approximate string matching there is also fuzzy_join(). All are S3 methods on vectra_node, share the same by syntax for key specification, and compose freely with other verbs in a pipeline.

We will work with ecological data throughout: species observations, site metadata, and a taxonomic reference table. Let us set those up.

library(vectra)
#> 
#> Attaching package: 'vectra'
#> The following object is masked from 'package:stats':
#> 
#>     filter

obs_path <- tempfile(fileext = ".vtr")
sites_path <- tempfile(fileext = ".vtr")
ref_path <- tempfile(fileext = ".vtr")

observations <- data.frame(
  site_id = c(1, 1, 2, 2, 3, 3, 4, 4),
  species = c("Quercus robur", "Pinus sylvestris",
              "Quercus robur", "Fagus sylvatica",
              "Pinus sylvestris", "Betula pendula",
              "Fagus sylvatica", "Acer pseudoplatanus"),
  date = c("2024-06-01", "2024-06-01", "2024-06-15",
           "2024-06-15", "2024-07-01", "2024-07-01",
           "2024-07-10", "2024-07-10"),
  count = c(12, 5, 8, 3, 7, 14, 6, 2),
  stringsAsFactors = FALSE
)

sites <- data.frame(
  site_id = c(1, 2, 3, 5),
  latitude = c(48.21, 47.07, 46.62, 48.85),
  longitude = c(16.37, 15.44, 14.31, 13.02),
  habitat = c("forest", "wetland", "grassland", "forest"),
  stringsAsFactors = FALSE
)

reference <- data.frame(
  species_name = c("Quercus robur", "Pinus sylvestris",
                   "Fagus sylvatica", "Betula pendula",
                   "Alnus glutinosa"),
  family = c("Fagaceae", "Pinaceae", "Fagaceae",
             "Betulaceae", "Betulaceae"),
  conservation_status = c("LC", "LC", "LC", "LC", "LC"),
  stringsAsFactors = FALSE
)

write_vtr(observations, obs_path)
write_vtr(sites, sites_path)
write_vtr(reference, ref_path)

Key specification

Every join except cross_join() takes a by argument that defines which columns to match on. There are four ways to specify keys.

Same-name column. The simplest case: both tables share a column name, and we pass it as a string.

left_join(tbl(obs_path), tbl(sites_path), by = "site_id") |>
  collect()
#>   site_id             species       date count latitude longitude   habitat
#> 1       1       Quercus robur 2024-06-01    12    48.21     16.37    forest
#> 2       1    Pinus sylvestris 2024-06-01     5    48.21     16.37    forest
#> 3       2       Quercus robur 2024-06-15     8    47.07     15.44   wetland
#> 4       2     Fagus sylvatica 2024-06-15     3    47.07     15.44   wetland
#> 5       3    Pinus sylvestris 2024-07-01     7    46.62     14.31 grassland
#> 6       3      Betula pendula 2024-07-01    14    46.62     14.31 grassland
#> 7       4     Fagus sylvatica 2024-07-10     6       NA        NA      <NA>
#> 8       4 Acer pseudoplatanus 2024-07-10     2       NA        NA      <NA>

The result contains all columns from both tables. The key column site_id appears once; the remaining columns from each side sit beside it.

Different-name columns. When the key has a different name in each table, we use a named character vector. The name is the left-side column; the value is the right-side column.

inner_join(
  tbl(obs_path),
  tbl(ref_path),
  by = c("species" = "species_name")
) |> collect()
#>   site_id          species       date count     family conservation_status
#> 1       1    Quercus robur 2024-06-01    12   Fagaceae                  LC
#> 2       1 Pinus sylvestris 2024-06-01     5   Pinaceae                  LC
#> 3       2    Quercus robur 2024-06-15     8   Fagaceae                  LC
#> 4       2  Fagus sylvatica 2024-06-15     3   Fagaceae                  LC
#> 5       3 Pinus sylvestris 2024-07-01     7   Pinaceae                  LC
#> 6       3   Betula pendula 2024-07-01    14 Betulaceae                  LC
#> 7       4  Fagus sylvatica 2024-07-10     6   Fagaceae                  LC

Here species in the observations table maps to species_name in the reference table. The output keeps the left-side name (species).

Natural join. When by is NULL (the default), vectra identifies all columns that appear in both tables and joins on all of them. A message tells us which columns were selected.

left_join(tbl(obs_path), tbl(sites_path)) |>
  collect()
#> Joining by: site_id
#>   site_id             species       date count latitude longitude   habitat
#> 1       1       Quercus robur 2024-06-01    12    48.21     16.37    forest
#> 2       1    Pinus sylvestris 2024-06-01     5    48.21     16.37    forest
#> 3       2       Quercus robur 2024-06-15     8    47.07     15.44   wetland
#> 4       2     Fagus sylvatica 2024-06-15     3    47.07     15.44   wetland
#> 5       3    Pinus sylvestris 2024-07-01     7    46.62     14.31 grassland
#> 6       3      Betula pendula 2024-07-01    14    46.62     14.31 grassland
#> 7       4     Fagus sylvatica 2024-07-10     6       NA        NA      <NA>
#> 8       4 Acer pseudoplatanus 2024-07-10     2       NA        NA      <NA>

Natural joins are convenient for exploratory work. For production pipelines, explicit keys are safer because they survive schema changes in upstream tables.

Multi-column keys. We pass multiple column names as a character vector. This defines a composite key where rows must match on every column simultaneously.

survey1_path <- tempfile(fileext = ".vtr")
survey2_path <- tempfile(fileext = ".vtr")

write_vtr(data.frame(
  site_id = c(1, 1, 2, 2),
  year = c(2023, 2024, 2023, 2024),
  richness = c(5, 7, 3, 4)
), survey1_path)

write_vtr(data.frame(
  site_id = c(1, 2, 2),
  year = c(2024, 2023, 2024),
  temperature = c(18.2, 16.1, 17.5)
), survey2_path)

inner_join(
  tbl(survey1_path), tbl(survey2_path),
  by = c("site_id", "year")
) |> collect()
#>   site_id year richness temperature
#> 1       1 2024        7        18.2
#> 2       2 2023        3        16.1
#> 3       2 2024        4        17.5

Only the three rows where both site_id and year match appear. Site 1 in 2023 had a richness measurement but no temperature record, so it drops from the inner join. Multi-column keys can also use the named form when column names differ: by = c("site" = "site_id", "yr" = "year").

Left join

left_join() keeps every row from the left table. Where a match exists in the right table, the right-side columns are filled in. Where no match exists, they become NA.

This is the workhorse join for data enrichment. We have a large observation table and want to attach site coordinates to each row.

enriched <- left_join(
  tbl(obs_path), tbl(sites_path), by = "site_id"
) |> collect()
enriched
#>   site_id             species       date count latitude longitude   habitat
#> 1       1       Quercus robur 2024-06-01    12    48.21     16.37    forest
#> 2       1    Pinus sylvestris 2024-06-01     5    48.21     16.37    forest
#> 3       2       Quercus robur 2024-06-15     8    47.07     15.44   wetland
#> 4       2     Fagus sylvatica 2024-06-15     3    47.07     15.44   wetland
#> 5       3    Pinus sylvestris 2024-07-01     7    46.62     14.31 grassland
#> 6       3      Betula pendula 2024-07-01    14    46.62     14.31 grassland
#> 7       4     Fagus sylvatica 2024-07-10     6       NA        NA      <NA>
#> 8       4 Acer pseudoplatanus 2024-07-10     2       NA        NA      <NA>

Site 4 has observations but does not appear in the sites table. Its latitude, longitude, and habitat columns are NA. Site 5 is in the sites table but has no observations; it does not appear at all because left join preserves the left side, not the right.

Internally, the engine builds a hash table from the right-side table (sites) during the build phase. Every row in sites is hashed on site_id and stored with its payload columns (latitude, longitude, habitat). Then the left side (observations) streams through one batch at a time. For each left row, the engine computes the hash of site_id, probes the hash table, and either copies the matching right-side columns into the output or fills them with NA if the probe misses. The left row is emitted regardless. This is why left join guarantees that the output has at least as many rows as the left table: every left row produces exactly one output row per match, or one NA-padded row if there is no match.

Left join is the natural default when we have a primary dataset and want to enrich it with attributes from a secondary source. Inner join, by contrast, drops every row that lacks a match on the right side, which can silently shrink the result. In our example, inner join would discard both observations at site 4. Whether that is acceptable depends on the analysis. If we plan to model species richness as a function of habitat, losing site 4 might be fine because we cannot use those rows without habitat data anyway. But if we are computing total species counts across all sites, dropping site 4 would undercount. Left join keeps the data intact and lets us decide later what to do with the gaps.

Those NA values in the output carry meaning: they tell us which observations lack metadata. We can filter them out, impute them, or flag them for data collection. A common pattern is to follow a left join with filter(!is.na(habitat)) when we want inner-join semantics but also want to inspect the unmatched rows first. Running the left join, checking the NA rate, and then filtering is a safer workflow than jumping straight to inner join because it makes the data loss visible.

The row count of a left join is always at least the row count of the left table. It can be larger if the right table has duplicate keys. If site_id 1 appeared twice in the sites table, each observation at site 1 would produce two output rows. A one-to-one key relationship keeps the row count equal; a one-to-many relationship on the right side inflates it. Checking nrow(result) == nrow(left) after a left join is a quick way to confirm the key relationship is one-to-one, which is often what we assume but rarely verify.

Suffix for duplicate column names. When both tables share a non-key column name, vectra appends suffixes to disambiguate. The default is c(".x", ".y").

extra_path <- tempfile(fileext = ".vtr")
write_vtr(data.frame(
  site_id = c(1, 2, 3),
  count = c(100, 200, 300)
), extra_path)

left_join(
  tbl(obs_path), tbl(extra_path),
  by = "site_id", suffix = c("_obs", "_site")
) |> collect()
#>   site_id             species       date count_obs count_site
#> 1       1       Quercus robur 2024-06-01        12        100
#> 2       1    Pinus sylvestris 2024-06-01         5        100
#> 3       2       Quercus robur 2024-06-15         8        200
#> 4       2     Fagus sylvatica 2024-06-15         3        200
#> 5       3    Pinus sylvestris 2024-07-01         7        300
#> 6       3      Betula pendula 2024-07-01        14        300
#> 7       4     Fagus sylvatica 2024-07-10         6         NA
#> 8       4 Acer pseudoplatanus 2024-07-10         2         NA

Both tables have a count column. The left side becomes count_obs, the right side count_site. Choosing descriptive suffixes avoids the generic .x/.y that can be hard to interpret in downstream code.

Inner join

inner_join() returns only the rows that match on both sides. It is the intersection of the two key sets.

inner_join(
  tbl(obs_path), tbl(sites_path), by = "site_id"
) |> collect()
#>   site_id          species       date count latitude longitude   habitat
#> 1       1    Quercus robur 2024-06-01    12    48.21     16.37    forest
#> 2       1 Pinus sylvestris 2024-06-01     5    48.21     16.37    forest
#> 3       2    Quercus robur 2024-06-15     8    47.07     15.44   wetland
#> 4       2  Fagus sylvatica 2024-06-15     3    47.07     15.44   wetland
#> 5       3 Pinus sylvestris 2024-07-01     7    46.62     14.31 grassland
#> 6       3   Betula pendula 2024-07-01    14    46.62     14.31 grassland

Site 4 (in observations but not in sites) and site 5 (in sites but not in observations) both vanish. The result has six rows instead of the eight we started with.

Inner join is the right choice when unmatched rows are meaningless for the analysis. If we only want observations where we have spatial metadata, inner join gives us exactly that. Compared to a left join followed by filter(!is.na(latitude)), the inner join does it in a single hash probe pass with no intermediate NA rows to filter out.

When the key relationship is one-to-one on both sides, inner join and left join produce the same row count. The difference shows up when there are unmatched keys. Checking row counts after each join is a quick sanity test for whether the key relationship is what we expected.

Right and full joins

Right join. right_join() is the mirror of left_join(): it keeps every row from the right table and fills NA on the left side where there is no match. Internally vectra implements it by swapping the two inputs and reordering the output columns to match the expected layout.

right_join(
  tbl(obs_path), tbl(sites_path), by = "site_id"
) |> collect()
#>   site_id          species       date count latitude longitude   habitat
#> 1       1 Pinus sylvestris 2024-06-01     5    48.21     16.37    forest
#> 2       1    Quercus robur 2024-06-01    12    48.21     16.37    forest
#> 3       2  Fagus sylvatica 2024-06-15     3    47.07     15.44   wetland
#> 4       2    Quercus robur 2024-06-15     8    47.07     15.44   wetland
#> 5       3   Betula pendula 2024-07-01    14    46.62     14.31 grassland
#> 6       3 Pinus sylvestris 2024-07-01     7    46.62     14.31 grassland
#> 7       5             <NA>       <NA>    NA    48.85     13.02    forest

Site 5 now appears with NA values for species, date, and count. Site 4 drops because it has no entry in the right table (sites). A right join with the tables in order (A, B) produces the same matches as a left join with (B, A), but the column order follows the original call: left-side columns first, then right-side columns.

Full join. full_join() keeps every row from both sides. Rows with a match get combined; unmatched rows from either side are padded with NA.

full_join(
  tbl(obs_path), tbl(sites_path), by = "site_id"
) |> collect()
#>   site_id             species       date count latitude longitude   habitat
#> 1       1       Quercus robur 2024-06-01    12    48.21     16.37    forest
#> 2       1    Pinus sylvestris 2024-06-01     5    48.21     16.37    forest
#> 3       2       Quercus robur 2024-06-15     8    47.07     15.44   wetland
#> 4       2     Fagus sylvatica 2024-06-15     3    47.07     15.44   wetland
#> 5       3    Pinus sylvestris 2024-07-01     7    46.62     14.31 grassland
#> 6       3      Betula pendula 2024-07-01    14    46.62     14.31 grassland
#> 7       4     Fagus sylvatica 2024-07-10     6       NA        NA      <NA>
#> 8       4 Acer pseudoplatanus 2024-07-10     2       NA        NA      <NA>
#> 9       5                <NA>       <NA>    NA    48.85     13.02    forest

Both site 4 (left-only) and site 5 (right-only) appear. Full join is useful when we need to see the complete picture and handle gaps explicitly downstream.

The engine processes full joins in three phases. First, it builds the hash table from the right side. Second, it streams the left side, emitting matched and unmatched left rows. Third, it walks the hash table to emit any right rows that were never probed, flushing them in 65536-row chunks. This third phase adds a memory overhead equal to the unmatched portion of the right side.

Filtering joins: semi and anti

Filtering joins do not bring in columns from the right table. They use the right table purely as a lookup to decide which left rows to keep or discard.

Semi join. semi_join() answers: “which of my left rows have at least one match in the right table?” The output schema is identical to the left table.

Consider a scenario where we have a list of target species and want to keep only observations of those species.

targets_path <- tempfile(fileext = ".vtr")
write_vtr(data.frame(
  species = c("Quercus robur", "Betula pendula"),
  stringsAsFactors = FALSE
), targets_path)
semi_join(
  tbl(obs_path), tbl(targets_path), by = "species"
) |> collect()
#>   site_id        species       date count
#> 1       1  Quercus robur 2024-06-01    12
#> 2       2  Quercus robur 2024-06-15     8
#> 3       3 Betula pendula 2024-07-01    14

Three observations match. The result has the same columns as the observations table; no columns from the target list leak through. If a species appeared multiple times in the targets table, each observation would still appear only once. Semi join deduplicates by design: the right table is a set, not a mapping.

Semi join shows up frequently in spatial and temporal subsetting. Suppose we have a table of sites that passed a quality control check, and we want to restrict a large observation dataset to only those sites. An inner join would work, but it would also attach the QC columns to every output row, widening the result for no reason. Semi join keeps the observation schema untouched. In pipelines where the output feeds into another join or a group-by aggregation, avoiding unnecessary columns reduces both memory use and cognitive load when reading the schema.

Anti join. anti_join() is the complement: “which left rows have no match in the right table?” This is the diagnostic workhorse for finding orphan records, missing reference entries, and data quality gaps.

The classic alternative to anti join is left_join() |> filter(is.na(key)). Both produce the same rows, but anti join is more efficient: the engine never copies right-side columns into the output, and it never produces intermediate NA-padded rows that the filter then discards. For a left table with 10 million rows and a right table with 50,000 rows, the left-join-then-filter approach allocates space for right-side payload columns in every output batch, fills most of them with real data, and then throws those rows away. Anti join skips all of that. It probes the hash table, checks for a miss, and emits the left row as-is.

Beyond performance, anti join communicates intent. A reader seeing anti_join() immediately knows the goal is to find what is missing. A left_join() followed by filter(is.na(...)) requires reading two lines and mentally combining them.

Which observed species are missing from our taxonomic reference list?

anti_join(
  tbl(obs_path),
  tbl(ref_path),
  by = c("species" = "species_name")
) |> collect()
#>   site_id             species       date count
#> 1       4 Acer pseudoplatanus 2024-07-10     2

Acer pseudoplatanus is in the observations but not in the reference table. In a real workflow, this flags an incomplete reference list or a data entry error.

The reverse question is equally useful. Which reference species have never been observed?

anti_join(
  tbl(ref_path),
  tbl(obs_path),
  by = c("species_name" = "species")
) |> collect()
#>      species_name     family conservation_status
#> 1 Alnus glutinosa Betulaceae                  LC

Alnus glutinosa is in the reference but absent from observations. Running anti joins in both directions is a standard check when validating relational integrity between tables.

Filtering joins are cheaper than mutating joins when all we need is the filter effect. The hash table is built the same way, but no columns are copied from the right side, and the output batch can reuse the left batch’s memory directly. For wide right-side tables with many columns, the savings can be substantial: a mutating join copies every payload column for every matched row, while a filtering join touches none of them.

Cross join

cross_join() produces the Cartesian product of two tables. Every row on the left is paired with every row on the right. There is no by parameter.

A common use is generating a complete grid of combinations. Suppose we want every site-year pair for three years of monitoring.

years_path <- tempfile(fileext = ".vtr")
write_vtr(data.frame(year = c(2022, 2023, 2024)), years_path)

grid <- cross_join(tbl(sites_path), tbl(years_path))
grid
#>    site_id latitude longitude   habitat year
#> 1        1    48.21     16.37    forest 2022
#> 2        1    48.21     16.37    forest 2023
#> 3        1    48.21     16.37    forest 2024
#> 4        2    47.07     15.44   wetland 2022
#> 5        2    47.07     15.44   wetland 2023
#> 6        2    47.07     15.44   wetland 2024
#> 7        3    46.62     14.31 grassland 2022
#> 8        3    46.62     14.31 grassland 2023
#> 9        3    46.62     14.31 grassland 2024
#> 10       5    48.85     13.02    forest 2022
#> 11       5    48.85     13.02    forest 2023
#> 12       5    48.85     13.02    forest 2024

The result has 4 * 3 = 12 rows (4 sites times 3 years). From here we could left join actual survey data onto this grid to identify which site-year combinations lack records.

Cross joins scale as O(n * m), so both tables must be small enough for the output to fit in memory. vectra collects both sides before expanding. Use this for generating lookup grids, parameter sweeps, or small combinatorial tables. For large tables it is almost never what we want.

Fuzzy joins

Ecological datasets are full of messy names. Field observers write “Quercus robar” instead of “Quercus robur”, or “Pinus sylvestris L.” instead of “Pinus sylvestris”. Exact joins miss these. fuzzy_join() matches rows by approximate string distance.

messy_path <- tempfile(fileext = ".vtr")
clean_path <- tempfile(fileext = ".vtr")

write_vtr(data.frame(
  obs_id = c(1, 2, 3, 4, 5),
  name = c("Quercus robar", "Pinus sylvestris L.",
           "Fagus silvatica", "Betula pendla",
           "Acer pseudoplatanus"),
  stringsAsFactors = FALSE
), messy_path)

write_vtr(data.frame(
  species_name = c("Quercus robur", "Pinus sylvestris",
                   "Fagus sylvatica", "Betula pendula",
                   "Acer pseudoplatanus"),
  family = c("Fagaceae", "Pinaceae", "Fagaceae",
             "Betulaceae", "Sapindaceae"),
  stringsAsFactors = FALSE
), clean_path)
fuzzy_join(
  tbl(messy_path),
  tbl(clean_path),
  by = c("name" = "species_name"),
  method = "dl",
  max_dist = 0.2
) |> collect()
#>   obs_id                name        species_name      family fuzzy_dist
#> 1      1       Quercus robar       Quercus robur    Fagaceae 0.07692308
#> 2      2 Pinus sylvestris L.    Pinus sylvestris    Pinaceae 0.15789474
#> 3      3     Fagus silvatica     Fagus sylvatica    Fagaceae 0.06666667
#> 4      4       Betula pendla      Betula pendula  Betulaceae 0.07142857
#> 5      5 Acer pseudoplatanus Acer pseudoplatanus Sapindaceae 0.00000000

The output includes a fuzzy_dist column with the normalized distance between the matched strings. Rows that exceed max_dist are excluded.

Choosing the right max_dist threshold requires balancing recall against precision. A threshold of 0.1 catches only very minor typos (one character in a ten-character string). A threshold of 0.3 catches more variation but starts admitting false matches: “Pinus mugo” and “Pinus nigra” differ by only a few characters and could match if the threshold is too loose. The cost of a false match in a species lookup is a wrong family or conservation status propagating silently through the analysis. Start with a conservative threshold (0.1 or 0.15), inspect the fuzzy_dist column in the output, and widen only if legitimate matches are being missed. Sorting the result by fuzzy_dist in descending order is a quick way to audit the worst matches and decide whether the threshold needs tightening.

The three available methods have different characteristics:

  • "dl" (Damerau-Levenshtein, the default): counts insertions, deletions, substitutions, and transpositions. Good general-purpose metric for typos.

  • "levenshtein": same as DL but without transpositions. Slightly stricter.

  • "jw" (Jaro-Winkler): gives more weight to matching prefixes. Designed for short strings like personal names; works well for genus-level matching.

Blocking for performance. Fuzzy joins compare every left row against every right row (within the distance threshold). For large tables this is expensive. The block_by parameter restricts comparisons to rows that share an exact match on a blocking column.

messy2_path <- tempfile(fileext = ".vtr")
clean2_path <- tempfile(fileext = ".vtr")

write_vtr(data.frame(
  genus = c("Quercus", "Pinus", "Fagus", "Betula"),
  name = c("Quercus robar", "Pinus sylvestris L.",
           "Fagus silvatica", "Betula pendla"),
  stringsAsFactors = FALSE
), messy2_path)

write_vtr(data.frame(
  genus = c("Quercus", "Pinus", "Fagus",
            "Betula", "Alnus"),
  species_name = c("Quercus robur", "Pinus sylvestris",
                   "Fagus sylvatica", "Betula pendula",
                   "Alnus glutinosa"),
  stringsAsFactors = FALSE
), clean2_path)
fuzzy_join(
  tbl(messy2_path),
  tbl(clean2_path),
  by = c("name" = "species_name"),
  method = "dl",
  max_dist = 0.25,
  block_by = c("genus" = "genus")
) |> collect()
#>     genus                name genus.y     species_name fuzzy_dist
#> 1 Quercus       Quercus robar Quercus    Quercus robur 0.07692308
#> 2   Pinus Pinus sylvestris L.   Pinus Pinus sylvestris 0.15789474
#> 3   Fagus     Fagus silvatica   Fagus  Fagus sylvatica 0.06666667
#> 4  Betula       Betula pendla  Betula   Betula pendula 0.07142857

With blocking, each Quercus name is only compared against other Quercus entries, each Pinus name against other Pinus entries, and so on. Without blocking, a fuzzy join between a left table with N rows and a right table with M rows computes N * M distance comparisons. Each comparison involves walking both strings character by character, so the total cost is O(N * M * L) where L is the average string length. For N = 10,000 and M = 100,000, that is a billion comparisons. Blocking partitions both tables by the blocking column and runs the fuzzy comparison only within matching partitions. If the blocking column splits the data into k roughly equal groups, the comparison count drops to N * M / k. With 200 genera, that billion becomes five million.

The blocking column must be an exact-match column that is already clean on both sides. Genus is a natural choice for species names because it is shorter, less prone to typos, and partitions the data well. Other good blocking columns include country codes, first letters of names, or any categorical variable that coarsely groups the data. If the blocking column itself has errors, rows will land in the wrong partition and miss their true match entirely, so it is worth cleaning the blocking column before relying on it.

The fuzzy distance computation within each block is parallelized with OpenMP, and the n_threads parameter controls the thread count.

Multi-column keys

When a single column does not uniquely identify the relationship, we join on a composite key. This is common with spatial and temporal data where the natural key is a site-year or site-date combination.

counts_path <- tempfile(fileext = ".vtr")
effort_path <- tempfile(fileext = ".vtr")

write_vtr(data.frame(
  site_id = c(1, 1, 2, 2, 3),
  year = c(2023, 2024, 2023, 2024, 2024),
  n_species = c(12, 15, 8, 10, 20)
), counts_path)

write_vtr(data.frame(
  site_id = c(1, 1, 2, 3, 3),
  year = c(2023, 2024, 2024, 2023, 2024),
  hours = c(4.5, 5.0, 3.0, 6.0, 4.0)
), effort_path)
left_join(
  tbl(counts_path), tbl(effort_path),
  by = c("site_id", "year")
) |> collect()
#>   site_id year n_species hours
#> 1       1 2023        12   4.5
#> 2       1 2024        15   5.0
#> 3       2 2024        10   3.0
#> 4       3 2024        20   4.0
#> 5       2 2023         8    NA

Site 2 in 2023 has a species count but no matching effort record, so the effort column is NA for that row. Site 3 in 2023 has effort data but no count, and because this is a left join it does not appear in the output at all.

Composite keys work with all join types. The hash table uses a combined hash of all key columns (FNV-1a hash combining), so lookup cost does not increase linearly with the number of key columns. There is a single hash probe per row regardless of whether the key has one column or five.

When column names differ between tables, the named form works for composites too:

effort2_path <- tempfile(fileext = ".vtr")
write_vtr(data.frame(
  loc = c(1, 1, 2, 3, 3),
  survey_year = c(2023, 2024, 2024, 2023, 2024),
  hours = c(4.5, 5.0, 3.0, 6.0, 4.0)
), effort2_path)

left_join(
  tbl(counts_path), tbl(effort2_path),
  by = c("site_id" = "loc", "year" = "survey_year")
) |> collect()
#>   site_id year n_species hours
#> 1       1 2023        12   4.5
#> 2       1 2024        15   5.0
#> 3       2 2024        10   3.0
#> 4       3 2024        20   4.0
#> 5       2 2023         8    NA

Key coercion and NA handling

vectra’s join keys follow a type hierarchy: bool < int64 < double. When the left key is int64 and the right key is double, or vice versa, both sides are coerced to double before hashing. This happens transparently.

int_path <- tempfile(fileext = ".vtr")
dbl_path <- tempfile(fileext = ".vtr")

write_vtr(data.frame(id = c(1L, 2L, 3L), x = c(10, 20, 30)), int_path)
write_vtr(data.frame(id = c(1.0, 2.0, 4.0), y = c(100, 200, 400)), dbl_path)

inner_join(tbl(int_path), tbl(dbl_path), by = "id") |> collect()
#>   id  x   y
#> 1  1 10 100
#> 2  2 20 200

The integer 1L matches the double 1.0 after coercion. Boolean columns coerce to int64 (FALSE = 0, TRUE = 1) if the other side is numeric.

Joining a string column against a numeric column is an error. vectra does not attempt to parse strings as numbers, because that coercion is ambiguous (“01” vs. 1, leading zeros, locale-dependent formatting). If we need to join string IDs against numeric IDs, we convert explicitly with mutate() before the join.

NA keys never match. This follows SQL NULL semantics. A row with NA in the key column will not match any other row, including another NA.

na_left <- tempfile(fileext = ".vtr")
na_right <- tempfile(fileext = ".vtr")

write_vtr(data.frame(
  id = c(1, NA, 3), x = c(10, 20, 30)
), na_left)
write_vtr(data.frame(
  id = c(1, NA, 4), y = c(100, 200, 400)
), na_right)

left_join(tbl(na_left), tbl(na_right), by = "id") |> collect()
#>   id  x   y
#> 1  1 10 100
#> 2 NA 20  NA
#> 3  3 30  NA

Row 2 on the left has id = NA. Even though row 2 on the right also has id = NA, they do not match. The left row is preserved (because this is a left join) but with NA for column y. In an inner join, both NA-keyed rows would drop entirely.

This behavior is consistent and predictable, but it catches people off guard when they expect NA-to-NA matching. The consequence for left joins is that a left row with an NA key will always appear in the output with NA-filled right columns, even if the right table also has rows with NA keys. Those right-side NA-keyed rows sit unused in the hash table. For inner joins, NA keys on either side mean those rows vanish from the output entirely. For anti joins, every NA-keyed left row is guaranteed to survive because it can never find a match.

In practice, NA keys usually signal missing data rather than a deliberate category. A site_id of NA means “we don’t know which site this observation belongs to,” not “this observation belongs to the NA site.” SQL NULL semantics align with that interpretation: unknown values cannot be equal. If matching on missing values is genuinely the goal (rare, but it happens when NA represents a shared “unknown” category), replace NA with a sentinel value via mutate() before joining.

Memory and performance

Understanding the memory profile of joins helps with sizing. vectra’s hash join has three cost components:

  1. Build phase. The entire right-side table is collected into a hash table. Memory cost is proportional to nrow(right) * avg_row_width. Each row in the hash table stores all columns (keys + payload), plus hash bucket pointers and overflow chains.

  2. Probe phase. The left side streams batch by batch. Only one batch lives in memory at a time, regardless of how large the left table is. For a table with 100 million rows and 10,000 rows per batch, only 10,000 rows of the left side are in memory at any given moment.

  3. Finalize phase (full join only). After probing, the engine walks the hash table to find unmatched right rows. These are flushed in 65536-row chunks. Memory overhead is small since the hash table already exists.

The practical rule: put the smaller table on the right. If we join a 500 million row observation table against a 10,000 row lookup table, the lookup table goes on the right. Memory usage is then roughly 10,000 * row_width bytes plus overhead for the hash table structure. The 500 million row left side streams through without ever being fully in memory.

To put concrete numbers on “right-side memory,” consider a typical lookup table with 5 columns: one int64 key, two double-precision floats, and two short strings (average 20 bytes each). Each row occupies roughly 8 + 8 + 8 + 20 + 20 = 64 bytes of payload, plus ~24 bytes of hash table overhead per slot (hash value, next-pointer, validity flags). That gives us roughly 90 bytes per row. A right table with 100,000 rows costs about 9 MB; one with 1 million rows costs about 90 MB. String-heavy tables cost more because each string value is stored inline. A table with a 200-character text column per row could easily triple the per-row cost.

When both tables are large and neither fits comfortably, pre-filtering the right side before the join is the single most effective optimization. A select() that drops columns not needed downstream reduces row width directly. A filter() that removes irrelevant rows reduces row count. Both shrink the hash table. If the right side has many duplicate keys and we only need one row per key (say, the most recent record), aggregating with group_by() |> summarise() or deduplicating with a window function before the join can cut the right-side size by an order of magnitude.

Filtering joins are cheaper than mutating joins when all we need is the existence check. semi_join() and anti_join() still build the right-side hash table, but they skip copying right-side payload columns into the output. The output reuses the left batch’s memory layout. This makes them faster and more memory-efficient than the equivalent pattern of left_join() followed by filter(is.na(...)) or filter(!is.na(...)).

Hash table sizing. The engine uses an open-addressing hash table with a 70% load factor threshold. For a right table with N rows, the table allocates roughly N / 0.7 slots. The hash function is FNV-1a, which distributes well for both numeric and string keys. Composite keys combine per-column hashes with FNV-1a mixing, so multi-column keys have the same probe cost as single column keys.

Practical guidance

Choosing the right join type is often the first design decision in a pipeline. Here is a decision framework.

Data enrichment: left join. We have a primary table and want to attach attributes from a lookup table. Not every row will match, and we want to keep all primary rows. This is the most common pattern in analytical work.

# Attach habitat type to every observation
left_join(tbl(obs_path), tbl(sites_path), by = "site_id") |>
  select(site_id, species, habitat) |>
  collect()
#>   site_id             species   habitat
#> 1       1       Quercus robur    forest
#> 2       1    Pinus sylvestris    forest
#> 3       2       Quercus robur   wetland
#> 4       2     Fagus sylvatica   wetland
#> 5       3    Pinus sylvestris grassland
#> 6       3      Betula pendula grassland
#> 7       4     Fagus sylvatica      <NA>
#> 8       4 Acer pseudoplatanus      <NA>

Exact intersection: inner join. We only want rows where both sides have data. Useful for restricting analysis to a validated subset.

Completeness check: full join. We want to see everything and handle gaps explicitly. Good for reconciliation between two data sources.

Set membership: semi join. We want to filter a table to rows that exist in a reference set, without adding any columns. More efficient than inner join when we do not need the right-side columns.

Finding gaps: anti join. Orphan records, missing entries, incomplete coverage. Anti join is the diagnostic tool for relational integrity. Run it in both directions to get the full picture.

All combinations: cross join. Generating a grid of factor levels for downstream left joins. Use sparingly; output grows multiplicatively.

Approximate matching: fuzzy join. When keys have typos, inconsistent formatting, or minor spelling variations that prevent exact matching. Always try preprocessing first (trimming whitespace, case normalization) because exact joins are faster by orders of magnitude.

# Preprocessing before an exact join
tbl(obs_path) |>
  mutate(species_clean = trimws(tolower(species))) |>
  select(site_id, species_clean, count) |>
  collect()
#>   site_id       species_clean count
#> 1       1       quercus robur    12
#> 2       1    pinus sylvestris     5
#> 3       2       quercus robur     8
#> 4       2     fagus sylvatica     3
#> 5       3    pinus sylvestris     7
#> 6       3      betula pendula    14
#> 7       4     fagus sylvatica     6
#> 8       4 acer pseudoplatanus     2

If tolower() and trimws() bring the keys into alignment, an exact join is the better choice. Reserve fuzzy_join() for genuine spelling variation that normalization cannot resolve.

Many-to-many relationships. When both tables have duplicate keys, the join produces a row for every combination of matching rows. If the left table has 3 rows with id = 1 and the right table has 4 rows with id = 1, the output has 12 rows for that key. This multiplicative blowup can be surprising and, for large tables, dangerous: a key that appears 1,000 times on the left and 1,000 times on the right produces a million output rows for that single key value.

Many-to-many joins usually signal a modelling error. The most common causes are: duplicate rows in a lookup table that was supposed to have unique keys, a join key that is too coarse (joining on year when the correct key is site_id + year), or a table that represents a genuinely many-to-many relationship but needs an intermediate mapping table. To detect the problem, check for duplicate keys on each side before joining. A quick group_by(key) |> summarise(n = n()) |> filter(n > 1) on the right table reveals duplicates. If the relationship is genuinely many-to-many and the blowup is intended (e.g., generating all pairwise combinations within groups), a cross join within groups may be more explicit about what is happening.

Join ordering in multi-join pipelines. When a pipeline chains several joins, the order matters for both correctness and performance. Consider an observation table that needs site metadata, taxonomic classification, and sampling effort. We have three right-side tables to attach. Performance-wise, the most selective join should go first: if the taxonomic reference covers only 80% of the observed species and we are doing an inner join on taxonomy, placing that join first reduces the row count before the subsequent joins build their hash tables. Each downstream join then processes fewer left-side rows. For left joins, order does not affect row count (every left row survives), but it still affects width: each join adds columns, and wider rows cost more to hash and copy in subsequent joins. Attaching the narrowest lookup table first keeps intermediate rows slim.

Self-joins for deduplication. Anti-joining a table against itself (with adjusted keys or filters) can identify duplicate records. For instance, finding observations that share a site, species, and date but have different counts.