dev.kwill.datomic-backup.current-state-restore — dev.kwill/datomic-backup 1.0.10

Current state restore for Datomic databases.

Restores a database by copying schema and current datom state (no history) from a source database to a destination database.

Restore Strategies

Single-Pass (default, :two-pass? false)

Processes all attributes together, managing ref dependencies with a pending index. Best for databases with few ref attributes or simple dependency graphs.

Two-Pass (:two-pass? true)

Processes non-ref attributes first, then ref attributes second. Benefits:

• Eliminates most pending overhead (pass 1 has zero pending)
• Near 100% batch efficiency in pass 1
• Minimal pending in pass 2 (only circular refs)
• Expected speedup: 1.5-2.5x for databases with many ref attributes
• Pass 1 supports parallel transactions (via :tx-parallelism)

Performance Tuning

The restore process has these main performance knobs:

• :max-batch-size (default 500) - Datoms per transaction Higher = fewer transactions, faster overall

• :read-parallelism (default 20) - Parallel attribute reads Higher = faster reads

• :read-chunk (default 5000) - Datoms per read chunk Higher = fewer read calls

• :tx-parallelism (default 4) - Parallel transaction workers for Pass 1 Only used when :two-pass? true Higher = more concurrent transactions (limited by Datomic transactor throughput)

Performance Metrics

When :debug true, logs every 10 batches with these metrics:

• :pending-count - Datoms waiting due to ref dependencies High count → increase :max-batch-size or try :two-pass? true

• :last-tx-ms / :avg-tx-ms - Transaction timing High values → Datomic writes are slow, may be at limit

• :batch-efficiency - Percentage of batch actually written Low (< 80%) → increase :max-batch-size to resolve more refs

• :utilization-pct - Channel buffer utilization (every 10s) High (> 80%) → writes are bottleneck (good!) Low (< 20%) → reads are slow, increase :read-parallelism

At completion:

• :avg-tx-ms - Overall average transaction time
• :final-pending-count - Should be 0
• :duration-sec - Total time to read all datoms

-full-copy clj

(-full-copy {:keys [source-db schema-lookup dest-conn max-batch-size debug
                    read-parallelism read-chunk init-state attrs]})

<anom!! clj

(<anom!! ch)

add-tuple-attrs! clj

(add-tuple-attrs! {:keys [dest-conn tuple-schema]})

Adds tuple attributes to schema and establishes their composite values.

anom! clj

(anom! x)

batch-datoms! clj

(batch-datoms! datom-ch work-ch max-batch-size eid->schema max-bootstrap-tx)

Batches datoms from input channel into max-batch-size batches. Filters out schema and bootstrap datoms. Closes work-ch when input channel closes.

batch-datoms-partitioned! clj

(batch-datoms-partitioned! datom-ch
                           worker-channels
                           max-batch-size
                           eid->schema
                           max-bootstrap-tx
                           debug)

Batches datoms from input channel and routes to worker channels based on entity ID. Uses consistent hashing to ensure all datoms for the same entity go to the same worker. This prevents entity splitting when max-batch-size is smaller than an entity's datom count. Filters out schema and bootstrap datoms. Closes all worker channels when input channel closes.

collect-results! clj

(collect-results! result-ch debug init-state)

Collects results from workers and merges them into final state. Logs progress every 5 transactions when debug is enabled. Returns when result-ch is closed.

copy-schema! clj

(copy-schema! {:keys [dest-conn schema-lookup attrs]})

establish-composite-tuple! clj

(establish-composite-tuple! conn {:keys [attr batch-size]})

Reasserts all values of attr, in batches of batch-size. This will establish values for any composite attributes built from attr.

monitored-chan! clj

(monitored-chan! ch
                 {:keys [runningf channel-name every-ms] :or {every-ms 10000}})

partition-attributes-by-ref clj

(partition-attributes-by-ref ident->schema)

Partitions attribute eids into :non-ref and :ref based on their value types.

Non-ref attributes include:

• All non-ref, non-tuple attributes
• Tuples that don't contain any ref types

Ref attributes include:

• Direct :db.type/ref attributes
• Tuples containing at least one ref type (composite, heterogeneous, or homogeneous)

partition-schema-by-deps clj

(partition-schema-by-deps source-schema)

pass1-parallel-copy clj

(pass1-parallel-copy {:keys [source-db dest-conn schema-lookup max-batch-size
                             debug read-parallelism read-chunk tx-parallelism
                             init-state attrs]})

Parallel transaction pipeline for Pass 1 (non-ref datoms).

Since Pass 1 datoms have no ref dependencies, we can transact batches in parallel without coordination between transactions.

Uses entity-partitioned batching to ensure all datoms for the same entity are routed to the same worker, preventing duplicate entity creation when max-batch-size splits an entity's datoms across multiple batches.

Pipeline: [Datom Reader] → [Partitioned Batcher] → [Worker 0 Ch] → [Worker 0] → [Worker 1 Ch] → [Worker 1] → [Worker N Ch] → [Worker N] → [Results Collector]

process-single-batch clj

(process-single-batch
  {:keys [old-id->new-id pending-index dest-conn eid->schema debug] :as acc}
  batch)

Process a single batch of datoms, updating the accumulator state.

read-datoms-in-parallel clj

(read-datoms-in-parallel source-db
                         {:keys [attrs dest-ch parallelism read-chunk debug]})

read-datoms-to-chan! clj

(read-datoms-to-chan! db argm dest-ch)

resolve-datom clj

(resolve-datom datom eid->schema old-id->new-id)

restore clj

(restore {:keys [source-db dest-conn tx-parallelism] :as argm})

Restores a database by copying schema and current datom state.

Options:

• :tx-parallelism - Number of parallel transaction workers for Pass 1 in two-pass mode.

restore-two-pass clj

(restore-two-pass {:keys [attrs schema-lookup tx-parallelism init-state]
                   :as argm})

Two-pass restore: non-ref datoms first, then ref datoms.

This approach eliminates most pending overhead by ensuring all entities exist before processing ref attributes.

Pass 1: Process all non-ref attributes

• Zero pending overhead (no ref dependencies)
• Can use parallel transactions

Pass 2: Process all ref attributes

• Most refs resolve immediately
• Always sequential (due to potential circular refs)

tempids->old-id-mapping clj

(tempids->old-id-mapping tempids allowed-ks)

Extracts old-id->new-id mapping from transaction result's :tempids. Tempid keys are strings representing old entity IDs.

thread-factory clj

(thread-factory name-prefix)

Creates a ThreadFactory that names threads with the given prefix and an incrementing counter.

tx-worker! clj

(tx-worker! work-ch result-ch dest-conn eid->schema debug worker-id)

Worker that processes batches from work-ch and sends results to result-ch. For Pass 1 non-ref datoms only - no pending logic needed. Maintains old-id->new-id state across batches to handle entities split across batches.

txify-datoms clj

(txify-datoms datoms pending-index eid->schema old-id->new-id)