Fix: bind parameters by position to fix BETWEEN and other duplicate-name binds

[suaron]

Summary

When a query reuses one column across more than one ? placeholder, the adapter sent only one value to the driver, so the driver rejected the prepared statement (it expected N values and got 1).

This affects any where with a range (Arel::Between), an IN array, or a raw SQL fragment that names the same column twice.

The fix stops duplicate column names from overwriting each other.

Reproduction

require "bundler/inline"

gemfile do
  source "https://rubygems.org"
  gem "activerecord", "~> 8.1"
  gem "activerecord-adbc-adapter", "0.0.1"
end

require "active_record"
require "activerecord-adbc-adapter"

ActiveRecord::Base.establish_connection(
  adapter: "adbc",
  driver:  "adbc_driver_sqlite",
  uri:     "file::memory:?cache=shared",
)

conn = ActiveRecord::Base.connection
conn.execute(<<~SQL)
  CREATE TABLE sales (
    id      INTEGER PRIMARY KEY,
    sold_on DATE,
    region  TEXT,
    amount  REAL
  )
SQL

conn.execute(<<~SQL)
  INSERT INTO sales (id, sold_on, region, amount) VALUES
    (1, '2025-01-15', 'US',   10.0),
    (2, '2025-02-15', 'EU',   20.0),
    (3, '2025-03-15', 'APAC', 30.0),
    (4, '2025-05-15', 'US',   40.0)
SQL

class Sale < ActiveRecord::Base
  self.table_name = "sales"
end

RULE = "─" * 60

def shorten_frame(frame)
  frame
    .sub(%r{.*/gems/}, "")
    .sub(%r{.*/lib/ruby/[^/]+/}, "")
end

def check(label, &block)
  result = block.call
  puts "✓ #{label} → #{result.inspect}"
  puts RULE
rescue => e
  puts "✗ #{label}"
  e.message.lines.each {|line| puts "↳ #{line.chomp}" }
  e.backtrace.first(3).each {|frame| puts " at #{shorten_frame(frame)}" }
  puts RULE
end

puts "running bind-binding checks against ADBC SQLite"

check("BETWEEN inclusive range") do
  Sale.where(sold_on: Date.new(2025, 1, 1)..Date.new(2025, 3, 31)).count
end

check("BETWEEN exclusive range") do
  Sale.where(sold_on: Date.new(2025, 1, 1)...Date.new(2025, 3, 31)).count
end

check("IN array") do
  Sale.where(id: [1, 2, 3]).count
end

check("raw SQL repeated placeholder") do
  Sale.where("region = ? OR region = ?", "US", "EU").count
end

check("range + raw fragment") do
  Sale.where(sold_on: Date.new(2025, 1, 1)..Date.new(2025, 12, 31))
      .where("amount < ?", 25.0)
      .count
end

Root cause

ActiveRecordADBCAdapter::DatabaseStatements#perform_query keyed the parameter RecordBatch by bind.name:

binds.zip(type_casted_binds) do |bind, type_casted_bind|
  # ... build Arrow array ...
  raw_records[bind.name] = array      # collides on duplicate name
end
record_batch = Arrow::RecordBatch.new(raw_records)
statement.bind(record_batch) { statement.execute[0] }

where(sold_on: from..to) emits two QueryAttribute binds, both named "sold_on", for:

WHERE "sales"."sold_on" BETWEEN ? AND ?

On the second iteration the hash assignment overwrites the first. The RecordBatch ends up with one column while the prepared statement still needs two values.

The fix

Key columns by position and append the bind name only for readability:

binds.zip(type_casted_binds).each_with_index do |(bind, type_casted_bind), i|
  # ... build Arrow array ...
  # Some binds (e.g. BETWEEN range values) aren't QueryAttributes and
  # have no #name. Position `i` is the real key; name just aids debugging.
  name = bind.respond_to?(:name) ? bind.name : "p"
  raw_records["#{i}_#{name}"] = array
end

The position prefix keeps keys unique (0_sold_on, 1_sold_on, ...) even when two binds share a name. The name suffix is only there to make the columns readable when debugging. Ruby preserves hash insertion order, so column order still matches the order of the ? placeholders.

Switching to binds.zip(type_casted_binds) also fixes a separate crash: raw SQL fragments like where("col >= ?", 5) arrive in binds as plain values, not QueryAttribute objects, so .name raised NoMethodError. The respond_to?(:name) guard falls back to "p", and the value comes from type_casted_binds, so the loop no longer cares about the bind's shape.

Test plan

New cases in test/test_model.rb:

• test_where_range_between — where(id: 1..2).count
• test_where_range_between_exclusive_end — where(id: 1...2).count
• test_where_raw_sql_fragment — where("id >= ?", 2)
• test_where_in_array — where(id: [1, 2, 3])
• test_where_mixed_range_and_raw — where(id: 1..2).where("id < ?", 2)
• test_limit — order(:id).limit(2) returns 2 rows
• test_offset — order(:id).limit(2).offset(2) returns the trailing row

CI runs the suite three times via ACTIVERECORD_ADBC_ADAPTER_BACKEND (sqlite default, postgresql, duckdb). All cases pass on each backend.

[kou]

+1

[kou]

Could you update the PR description to reflect the latest change? We'll use the PR description as a commit message.

[suaron]

Could you update the PR description to reflect the latest change? We'll use the PR description as a commit message.

PR description updated to reflect the latest change and tidied up for use as the commit message.

[kou]

Thanks.