bio2zarr 0.0.5__py3-none-any.whl → 0.0.9__py3-none-any.whl

bio2zarr/vcf.py

@@ -1,17 +1,17 @@
 import collections
 import contextlib
 import dataclasses
-import functools
 import json
 import logging
 import math
 import os
+import os.path
 import pathlib
 import pickle
 import shutil
 import sys
 import tempfile
-from typing import Any, List
+from typing import Any
 
 import cyvcf2
 import humanfriendly
@@ -111,9 +111,6 @@ class VcfField:
             return self.name
         return f"{self.category}/{self.name}"
 
-    # TODO add method here to choose a good set compressor and
-    # filters default here for this field.
-
     def smallest_dtype(self):
         """
         Returns the smallest dtype suitable for this field based
@@ -123,13 +120,13 @@ class VcfField:
         if self.vcf_type == "Float":
             ret = "f4"
         elif self.vcf_type == "Integer":
-            dtype = "i4"
-            for a_dtype in ["i1", "i2"]:
-                info = np.iinfo(a_dtype)
-                if info.min <= s.min_value and s.max_value <= info.max:
-                    dtype = a_dtype
-                    break
-            ret = dtype
+            if not math.isfinite(s.max_value):
+                # All missing values; use i1. Note we should have some API to
+                # check more explicitly for missingness:
+                # https://github.com/sgkit-dev/bio2zarr/issues/131
+                ret = "i1"
+            else:
+                ret = core.min_int_dtype(s.min_value, s.max_value)
         elif self.vcf_type == "Flag":
             ret = "bool"
         elif self.vcf_type == "Character":
@@ -147,25 +144,41 @@ class VcfPartition:
     num_records: int = -1
 
 
-ICF_METADATA_FORMAT_VERSION = "0.2"
+ICF_METADATA_FORMAT_VERSION = "0.3"
 ICF_DEFAULT_COMPRESSOR = numcodecs.Blosc(
     cname="zstd", clevel=7, shuffle=numcodecs.Blosc.NOSHUFFLE
 )
 
 
+@dataclasses.dataclass
+class Contig:
+    id: str
+    length: int = None
+
+
+@dataclasses.dataclass
+class Sample:
+    id: str
+
+
+@dataclasses.dataclass
+class Filter:
+    id: str
+    description: str = ""
+
+
 @dataclasses.dataclass
 class IcfMetadata:
     samples: list
-    contig_names: list
-    contig_record_counts: dict
+    contigs: list
     filters: list
     fields: list
     partitions: list = None
-    contig_lengths: list = None
     format_version: str = None
     compressor: dict = None
     column_chunk_size: int = None
     provenance: dict = None
+    num_records: int = -1
 
     @property
     def info_fields(self):
@@ -184,8 +197,12 @@ class IcfMetadata:
         return fields
 
     @property
-    def num_records(self):
-        return sum(self.contig_record_counts.values())
+    def num_contigs(self):
+        return len(self.contigs)
+
+    @property
+    def num_filters(self):
+        return len(self.filters)
 
     @staticmethod
     def fromdict(d):
@@ -194,18 +211,23 @@ class IcfMetadata:
                 "Intermediate columnar metadata format version mismatch: "
                 f"{d['format_version']} != {ICF_METADATA_FORMAT_VERSION}"
             )
-        fields = [VcfField.fromdict(fd) for fd in d["fields"]]
         partitions = [VcfPartition(**pd) for pd in d["partitions"]]
         for p in partitions:
             p.region = vcf_utils.Region(**p.region)
         d = d.copy()
-        d["fields"] = fields
         d["partitions"] = partitions
+        d["fields"] = [VcfField.fromdict(fd) for fd in d["fields"]]
+        d["samples"] = [Sample(**sd) for sd in d["samples"]]
+        d["filters"] = [Filter(**fd) for fd in d["filters"]]
+        d["contigs"] = [Contig(**cd) for cd in d["contigs"]]
         return IcfMetadata(**d)
 
     def asdict(self):
         return dataclasses.asdict(self)
 
+    def asjson(self):
+        return json.dumps(self.asdict(), indent=4)
+
 
 def fixed_vcf_field_definitions():
     def make_field_def(name, vcf_type, vcf_number):
@@ -233,15 +255,22 @@ def fixed_vcf_field_definitions():
 def scan_vcf(path, target_num_partitions):
     with vcf_utils.IndexedVcf(path) as indexed_vcf:
         vcf = indexed_vcf.vcf
-        filters = [
-            h["ID"]
-            for h in vcf.header_iter()
-            if h["HeaderType"] == "FILTER" and isinstance(h["ID"], str)
-        ]
+        filters = []
+        pass_index = -1
+        for h in vcf.header_iter():
+            if h["HeaderType"] == "FILTER" and isinstance(h["ID"], str):
+                try:
+                    description = h["Description"].strip('"')
+                except KeyError:
+                    description = ""
+                if h["ID"] == "PASS":
+                    pass_index = len(filters)
+                filters.append(Filter(h["ID"], description))
+
         # Ensure PASS is the first filter if present
-        if "PASS" in filters:
-            filters.remove("PASS")
-            filters.insert(0, "PASS")
+        if pass_index > 0:
+            pass_filter = filters.pop(pass_index)
+            filters.insert(0, pass_filter)
 
         fields = fixed_vcf_field_definitions()
         for h in vcf.header_iter():
@@ -252,18 +281,22 @@ def scan_vcf(path, target_num_partitions):
                     field.vcf_number = "."
                 fields.append(field)
 
+        try:
+            contig_lengths = vcf.seqlens
+        except AttributeError:
+            contig_lengths = [None for _ in vcf.seqnames]
+
         metadata = IcfMetadata(
-            samples=vcf.samples,
-            contig_names=vcf.seqnames,
-            contig_record_counts=indexed_vcf.contig_record_counts(),
+            samples=[Sample(sample_id) for sample_id in vcf.samples],
+            contigs=[
+                Contig(contig_id, length)
+                for contig_id, length in zip(vcf.seqnames, contig_lengths)
+            ],
             filters=filters,
             fields=fields,
             partitions=[],
+            num_records=sum(indexed_vcf.contig_record_counts().values()),
         )
-        try:
-            metadata.contig_lengths = vcf.seqlens
-        except AttributeError:
-            pass
 
         regions = indexed_vcf.partition_into_regions(num_parts=target_num_partitions)
         logger.info(
@@ -282,21 +315,6 @@ def scan_vcf(path, target_num_partitions):
         return metadata, vcf.raw_header
 
 
-def check_overlap(partitions):
-    for i in range(1, len(partitions)):
-        prev_partition = partitions[i - 1]
-        current_partition = partitions[i]
-        if (
-            prev_partition.region.contig == current_partition.region.contig
-            and prev_partition.region.end > current_partition.region.start
-        ):
-            raise ValueError(
-                f"Multiple VCFs have the region "
-                f"{prev_partition.region.contig}:{prev_partition.region.start}-"
-                f"{current_partition.region.end}"
-            )
-
-
 def scan_vcfs(paths, show_progress, target_num_partitions, worker_processes=1):
     logger.info(
         f"Scanning {len(paths)} VCFs attempting to split into {target_num_partitions}"
@@ -325,27 +343,30 @@ def scan_vcfs(paths, show_progress, target_num_partitions, worker_processes=1):
     # We just take the first header, assuming the others
     # are compatible.
     all_partitions = []
-    contig_record_counts = collections.Counter()
+    total_records = 0
     for metadata, _ in results:
-        all_partitions.extend(metadata.partitions)
-        metadata.partitions.clear()
-        contig_record_counts += metadata.contig_record_counts
-        metadata.contig_record_counts.clear()
+        for partition in metadata.partitions:
+            logger.debug(f"Scanned partition {partition}")
+            all_partitions.append(partition)
+        total_records += metadata.num_records
+        metadata.num_records = 0
+        metadata.partitions = []
 
     icf_metadata, header = results[0]
     for metadata, _ in results[1:]:
         if metadata != icf_metadata:
             raise ValueError("Incompatible VCF chunks")
 
-    icf_metadata.contig_record_counts = dict(contig_record_counts)
+    # Note: this will be infinity here if any of the chunks has an index
+    # that doesn't keep track of the number of records per-contig
+    icf_metadata.num_records = total_records
 
     # Sort by contig (in the order they appear in the header) first,
     # then by start coordinate
-    contig_index_map = {contig: j for j, contig in enumerate(metadata.contig_names)}
+    contig_index_map = {contig.id: j for j, contig in enumerate(metadata.contigs)}
     all_partitions.sort(
         key=lambda x: (contig_index_map[x.region.contig], x.region.start)
     )
-    check_overlap(all_partitions)
     icf_metadata.partitions = all_partitions
     logger.info(f"Scan complete, resulting in {len(all_partitions)} partitions.")
     return icf_metadata, header
@@ -443,7 +464,7 @@ def sanitise_value_float_2d(buff, j, value):
 
 def sanitise_int_array(value, ndmin, dtype):
     if isinstance(value, tuple):
-        value = [VCF_INT_MISSING if x is None else x for x in value]  #  NEEDS TEST
+        value = [VCF_INT_MISSING if x is None else x for x in value]  # NEEDS TEST
     value = np.array(value, ndmin=ndmin, copy=False)
     value[value == VCF_INT_MISSING] = -1
     value[value == VCF_INT_FILL] = -2
@@ -736,9 +757,9 @@ class IcfFieldWriter:
     transformer: VcfValueTransformer
     compressor: Any
     max_buffered_bytes: int
-    buff: List[Any] = dataclasses.field(default_factory=list)
+    buff: list[Any] = dataclasses.field(default_factory=list)
     buffered_bytes: int = 0
-    chunk_index: List[int] = dataclasses.field(default_factory=lambda: [0])
+    chunk_index: list[int] = dataclasses.field(default_factory=lambda: [0])
     num_records: int = 0
 
     def append(self, val):
@@ -842,19 +863,18 @@ class IntermediateColumnarFormat(collections.abc.Mapping):
             self.metadata = IcfMetadata.fromdict(json.load(f))
         with open(self.path / "header.txt") as f:
             self.vcf_header = f.read()
-
         self.compressor = numcodecs.get_codec(self.metadata.compressor)
-        self.columns = {}
+        self.fields = {}
         partition_num_records = [
             partition.num_records for partition in self.metadata.partitions
         ]
         # Allow us to find which partition a given record is in
         self.partition_record_index = np.cumsum([0, *partition_num_records])
         for field in self.metadata.fields:
-            self.columns[field.full_name] = IntermediateColumnarFormatField(self, field)
+            self.fields[field.full_name] = IntermediateColumnarFormatField(self, field)
         logger.info(
             f"Loaded IntermediateColumnarFormat(partitions={self.num_partitions}, "
-            f"records={self.num_records}, columns={self.num_columns})"
+            f"records={self.num_records}, fields={self.num_fields})"
         )
 
     def __repr__(self):
@@ -865,17 +885,17 @@ class IntermediateColumnarFormat(collections.abc.Mapping):
         )
 
     def __getitem__(self, key):
-        return self.columns[key]
+        return self.fields[key]
 
     def __iter__(self):
-        return iter(self.columns)
+        return iter(self.fields)
 
     def __len__(self):
-        return len(self.columns)
+        return len(self.fields)
 
     def summary_table(self):
         data = []
-        for name, col in self.columns.items():
+        for name, col in self.fields.items():
             summary = col.vcf_field.summary
             d = {
                 "name": name,
@@ -891,9 +911,9 @@ class IntermediateColumnarFormat(collections.abc.Mapping):
             data.append(d)
         return data
 
-    @functools.cached_property
+    @property
     def num_records(self):
-        return sum(self.metadata.contig_record_counts.values())
+        return self.metadata.num_records
 
     @property
     def num_partitions(self):
@@ -904,8 +924,42 @@ class IntermediateColumnarFormat(collections.abc.Mapping):
         return len(self.metadata.samples)
 
     @property
-    def num_columns(self):
-        return len(self.columns)
+    def num_fields(self):
+        return len(self.fields)
+
+
+@dataclasses.dataclass
+class IcfPartitionMetadata:
+    num_records: int
+    last_position: int
+    field_summaries: dict
+
+    def asdict(self):
+        return dataclasses.asdict(self)
+
+    def asjson(self):
+        return json.dumps(self.asdict(), indent=4)
+
+    @staticmethod
+    def fromdict(d):
+        md = IcfPartitionMetadata(**d)
+        for k, v in md.field_summaries.items():
+            md.field_summaries[k] = VcfFieldSummary.fromdict(v)
+        return md
+
+
+def check_overlapping_partitions(partitions):
+    for i in range(1, len(partitions)):
+        prev_region = partitions[i - 1].region
+        current_region = partitions[i].region
+        if prev_region.contig == current_region.contig:
+            assert prev_region.end is not None
+            # Regions are *inclusive*
+            if prev_region.end >= current_region.start:
+                raise ValueError(
+                    f"Overlapping VCF regions in partitions {i - 1} and {i}: "
+                    f"{prev_region} and {current_region}"
+                )
 
 
 class IntermediateColumnarFormatWriter:
@@ -979,11 +1033,8 @@ class IntermediateColumnarFormatWriter:
         not_found = []
         for j in range(self.num_partitions):
             try:
-                with open(self.wip_path / f"p{j}_summary.json") as f:
-                    summary = json.load(f)
-                    for k, v in summary["field_summaries"].items():
-                        summary["field_summaries"][k] = VcfFieldSummary.fromdict(v)
-                    summaries.append(summary)
+                with open(self.wip_path / f"p{j}.json") as f:
+                    summaries.append(IcfPartitionMetadata.fromdict(json.load(f)))
             except FileNotFoundError:
                 not_found.append(j)
         if len(not_found) > 0:
@@ -1000,7 +1051,7 @@ class IntermediateColumnarFormatWriter:
 
     def process_partition(self, partition_index):
         self.load_metadata()
-        summary_path = self.wip_path / f"p{partition_index}_summary.json"
+        summary_path = self.wip_path / f"p{partition_index}.json"
         # If someone is rewriting a summary path (for whatever reason), make sure it
         # doesn't look like it's already been completed.
         # NOTE to do this properly we probably need to take a lock on this file - but
@@ -1021,6 +1072,7 @@ class IntermediateColumnarFormatWriter:
             else:
                 format_fields.append(field)
 
+        last_position = None
         with IcfPartitionWriter(
             self.metadata,
             self.path,
@@ -1030,6 +1082,7 @@ class IntermediateColumnarFormatWriter:
                 num_records = 0
                 for variant in ivcf.variants(partition.region):
                     num_records += 1
+                    last_position = variant.POS
                     tcw.append("CHROM", variant.CHROM)
                     tcw.append("POS", variant.POS)
                     tcw.append("QUAL", variant.QUAL)
@@ -1054,37 +1107,32 @@ class IntermediateColumnarFormatWriter:
                 f"flushing buffers"
             )
 
-        partition_metadata = {
-            "num_records": num_records,
-            "field_summaries": {k: v.asdict() for k, v in tcw.field_summaries.items()},
-        }
+        partition_metadata = IcfPartitionMetadata(
+            num_records=num_records,
+            last_position=last_position,
+            field_summaries=tcw.field_summaries,
+        )
         with open(summary_path, "w") as f:
-            json.dump(partition_metadata, f, indent=4)
+            f.write(partition_metadata.asjson())
         logger.info(
-            f"Finish p{partition_index} {partition.vcf_path}__{partition.region}="
-            f"{num_records} records"
+            f"Finish p{partition_index} {partition.vcf_path}__{partition.region} "
+            f"{num_records} records last_pos={last_position}"
         )
 
-    def process_partition_slice(
-        self,
-        start,
-        stop,
-        *,
-        worker_processes=1,
-        show_progress=False,
-    ):
+    def explode(self, *, worker_processes=1, show_progress=False):
         self.load_metadata()
-        if start == 0 and stop == self.num_partitions:
-            num_records = self.metadata.num_records
-        else:
-            # We only know the number of records if all partitions are done at once,
-            # and we signal this to tqdm by passing None as the total.
+        num_records = self.metadata.num_records
+        if np.isinf(num_records):
+            logger.warning(
+                "Total records unknown, cannot show progress; "
+                "reindex VCFs with bcftools index to fix"
+            )
             num_records = None
-        num_columns = len(self.metadata.fields)
+        num_fields = len(self.metadata.fields)
         num_samples = len(self.metadata.samples)
         logger.info(
-            f"Exploding columns={num_columns} samples={num_samples}; "
-            f"partitions={stop - start} "
+            f"Exploding fields={num_fields} samples={num_samples}; "
+            f"partitions={self.num_partitions} "
             f"variants={'unknown' if num_records is None else num_records}"
         )
         progress_config = core.ProgressConfig(
@@ -1094,48 +1142,43 @@ class IntermediateColumnarFormatWriter:
             show=show_progress,
         )
         with core.ParallelWorkManager(worker_processes, progress_config) as pwm:
-            for j in range(start, stop):
+            for j in range(self.num_partitions):
                 pwm.submit(self.process_partition, j)
 
-    def explode(self, *, worker_processes=1, show_progress=False):
-        self.load_metadata()
-        return self.process_partition_slice(
-            0,
-            self.num_partitions,
-            worker_processes=worker_processes,
-            show_progress=show_progress,
-        )
-
-    def explode_partition(self, partition, *, show_progress=False, worker_processes=1):
+    def explode_partition(self, partition):
         self.load_metadata()
         if partition < 0 or partition >= self.num_partitions:
             raise ValueError(
                 "Partition index must be in the range 0 <= index < num_partitions"
             )
-        return self.process_partition_slice(
-            partition,
-            partition + 1,
-            worker_processes=worker_processes,
-            show_progress=show_progress,
-        )
+        self.process_partition(partition)
 
     def finalise(self):
         self.load_metadata()
         partition_summaries = self.load_partition_summaries()
         total_records = 0
         for index, summary in enumerate(partition_summaries):
-            partition_records = summary["num_records"]
+            partition_records = summary.num_records
             self.metadata.partitions[index].num_records = partition_records
+            self.metadata.partitions[index].region.end = summary.last_position
             total_records += partition_records
-        assert total_records == self.metadata.num_records
+        if not np.isinf(self.metadata.num_records):
+            # Note: this is just telling us that there's a bug in the
+            # index based record counting code, but it doesn't actually
+            # matter much. We may want to just make this a warning if
+            # we hit regular problems.
+            assert total_records == self.metadata.num_records
+        self.metadata.num_records = total_records
+
+        check_overlapping_partitions(self.metadata.partitions)
 
         for field in self.metadata.fields:
             for summary in partition_summaries:
-                field.summary.update(summary["field_summaries"][field.full_name])
+                field.summary.update(summary.field_summaries[field.full_name])
 
         logger.info("Finalising metadata")
         with open(self.path / "metadata.json", "w") as f:
-            json.dump(self.metadata.asdict(), f, indent=4)
+            f.write(self.metadata.asjson())
 
         logger.debug("Removing WIP directory")
         shutil.rmtree(self.wip_path)
@@ -1186,14 +1229,9 @@ def explode_init(
     )
 
 
-# NOTE only including worker_processes here so we can use the 0 option to get the
-# work done syncronously and so we can get test coverage on it. Should find a
-# better way to do this.
-def explode_partition(icf_path, partition, *, show_progress=False, worker_processes=1):
+def explode_partition(icf_path, partition):
     writer = IntermediateColumnarFormatWriter(icf_path)
-    writer.explode_partition(
-        partition, show_progress=show_progress, worker_processes=worker_processes
-    )
+    writer.explode_partition(partition)
 
 
 def explode_finalise(icf_path):
@@ -1242,6 +1280,50 @@ class ZarrColumnSpec:
         spec._choose_compressor_settings()
         return spec
 
+    @staticmethod
+    def from_field(
+        vcf_field,
+        *,
+        num_variants,
+        num_samples,
+        variants_chunk_size,
+        samples_chunk_size,
+        variable_name=None,
+    ):
+        shape = [num_variants]
+        prefix = "variant_"
+        dimensions = ["variants"]
+        chunks = [variants_chunk_size]
+        if vcf_field.category == "FORMAT":
+            prefix = "call_"
+            shape.append(num_samples)
+            chunks.append(samples_chunk_size)
+            dimensions.append("samples")
+        if variable_name is None:
+            variable_name = prefix + vcf_field.name
+        # TODO make an option to add in the empty extra dimension
+        if vcf_field.summary.max_number > 1:
+            shape.append(vcf_field.summary.max_number)
+            # TODO we should really be checking this to see if the named dimensions
+            # are actually correct.
+            if vcf_field.vcf_number == "R":
+                dimensions.append("alleles")
+            elif vcf_field.vcf_number == "A":
+                dimensions.append("alt_alleles")
+            elif vcf_field.vcf_number == "G":
+                dimensions.append("genotypes")
+            else:
+                dimensions.append(f"{vcf_field.category}_{vcf_field.name}_dim")
+        return ZarrColumnSpec.new(
+            vcf_field=vcf_field.full_name,
+            name=variable_name,
+            dtype=vcf_field.smallest_dtype(),
+            shape=shape,
+            chunks=chunks,
+            dimensions=dimensions,
+            description=vcf_field.description,
+        )
+
     def _choose_compressor_settings(self):
         """
         Choose compressor and filter settings based on the size and
@@ -1250,19 +1332,34 @@ class ZarrColumnSpec:
 
         See https://github.com/pystatgen/bio2zarr/discussions/74
         """
-        dt = np.dtype(self.dtype)
         # Default is to not shuffle, because autoshuffle isn't recognised
         # by many Zarr implementations, and shuffling can lead to worse
         # performance in some cases anyway. Turning on shuffle should be a
         # deliberate choice.
         shuffle = numcodecs.Blosc.NOSHUFFLE
-        if dt.itemsize == 1:
-            # Any 1 byte field gets BITSHUFFLE by default
+        if self.name == "call_genotype" and self.dtype == "i1":
+            # call_genotype gets BITSHUFFLE by default as it gets
+            # significantly better compression (at a cost of slower
+            # decoding)
+            shuffle = numcodecs.Blosc.BITSHUFFLE
+        elif self.dtype == "bool":
             shuffle = numcodecs.Blosc.BITSHUFFLE
+
         self.compressor["shuffle"] = shuffle
 
+    @property
+    def variant_chunk_nbytes(self):
+        """
+        Returns the nbytes for a single variant chunk of this array.
+        """
+        chunk_items = self.chunks[0]
+        for size in self.shape[1:]:
+            chunk_items *= size
+        dt = np.dtype(self.dtype)
+        return chunk_items * dt.itemsize
+
 
-ZARR_SCHEMA_FORMAT_VERSION = "0.2"
+ZARR_SCHEMA_FORMAT_VERSION = "0.3"
 
 
 @dataclasses.dataclass
@@ -1271,11 +1368,10 @@ class VcfZarrSchema:
     samples_chunk_size: int
     variants_chunk_size: int
     dimensions: list
-    sample_id: list
-    contig_id: list
-    contig_length: list
-    filter_id: list
-    columns: dict
+    samples: list
+    contigs: list
+    filters: list
+    fields: dict
 
     def asdict(self):
         return dataclasses.asdict(self)
@@ -1291,8 +1387,11 @@ class VcfZarrSchema:
                 f"{d['format_version']} != {ZARR_SCHEMA_FORMAT_VERSION}"
             )
         ret = VcfZarrSchema(**d)
-        ret.columns = {
-            key: ZarrColumnSpec(**value) for key, value in d["columns"].items()
+        ret.samples = [Sample(**sd) for sd in d["samples"]]
+        ret.contigs = [Contig(**sd) for sd in d["contigs"]]
+        ret.filters = [Filter(**sd) for sd in d["filters"]]
+        ret.fields = {
+            key: ZarrColumnSpec(**value) for key, value in d["fields"].items()
         }
         return ret
 
@@ -1313,6 +1412,16 @@ class VcfZarrSchema:
             f"Generating schema with chunks={variants_chunk_size, samples_chunk_size}"
         )
 
+        def spec_from_field(field, variable_name=None):
+            return ZarrColumnSpec.from_field(
+                field,
+                num_samples=n,
+                num_variants=m,
+                samples_chunk_size=samples_chunk_size,
+                variants_chunk_size=variants_chunk_size,
+                variable_name=variable_name,
+            )
+
         def fixed_field_spec(
             name, dtype, vcf_field=None, shape=(m,), dimensions=("variants",)
         ):
@@ -1326,97 +1435,58 @@ class VcfZarrSchema:
                 chunks=[variants_chunk_size],
             )
 
-        alt_col = icf.columns["ALT"]
+        alt_col = icf.fields["ALT"]
         max_alleles = alt_col.vcf_field.summary.max_number + 1
-        num_filters = len(icf.metadata.filters)
 
-        # # FIXME get dtype from lookup table
         colspecs = [
             fixed_field_spec(
                 name="variant_contig",
-                dtype="i2",  # FIXME
+                dtype=core.min_int_dtype(0, icf.metadata.num_contigs),
             ),
             fixed_field_spec(
                 name="variant_filter",
                 dtype="bool",
-                shape=(m, num_filters),
+                shape=(m, icf.metadata.num_filters),
                 dimensions=["variants", "filters"],
             ),
             fixed_field_spec(
                 name="variant_allele",
                 dtype="str",
-                shape=[m, max_alleles],
+                shape=(m, max_alleles),
                 dimensions=["variants", "alleles"],
             ),
             fixed_field_spec(
-                vcf_field="POS",
-                name="variant_position",
-                dtype="i4",
-            ),
-            fixed_field_spec(
-                vcf_field=None,
                 name="variant_id",
                 dtype="str",
             ),
             fixed_field_spec(
-                vcf_field=None,
                 name="variant_id_mask",
                 dtype="bool",
             ),
-            fixed_field_spec(
-                vcf_field="QUAL",
-                name="variant_quality",
-                dtype="f4",
-            ),
         ]
+        name_map = {field.full_name: field for field in icf.metadata.fields}
+
+        # Only two of the fixed fields have a direct one-to-one mapping.
+        colspecs.extend(
+            [
+                spec_from_field(name_map["QUAL"], variable_name="variant_quality"),
+                spec_from_field(name_map["POS"], variable_name="variant_position"),
+            ]
+        )
+        colspecs.extend([spec_from_field(field) for field in icf.metadata.info_fields])
 
         gt_field = None
-        for field in icf.metadata.fields:
-            if field.category == "fixed":
-                continue
+        for field in icf.metadata.format_fields:
             if field.name == "GT":
                 gt_field = field
                 continue
-            shape = [m]
-            prefix = "variant_"
-            dimensions = ["variants"]
-            chunks = [variants_chunk_size]
-            if field.category == "FORMAT":
-                prefix = "call_"
-                shape.append(n)
-                chunks.append(samples_chunk_size)
-                dimensions.append("samples")
-            # TODO make an option to add in the empty extra dimension
-            if field.summary.max_number > 1:
-                shape.append(field.summary.max_number)
-                # TODO we should really be checking this to see if the named dimensions
-                # are actually correct.
-                if field.vcf_number == "R":
-                    dimensions.append("alleles")
-                elif field.vcf_number == "A":
-                    dimensions.append("alt_alleles")
-                elif field.vcf_number == "G":
-                    dimensions.append("genotypes")
-                else:
-                    dimensions.append(f"{field.category}_{field.name}_dim")
-            variable_name = prefix + field.name
-            colspec = ZarrColumnSpec.new(
-                vcf_field=field.full_name,
-                name=variable_name,
-                dtype=field.smallest_dtype(),
-                shape=shape,
-                chunks=chunks,
-                dimensions=dimensions,
-                description=field.description,
-            )
-            colspecs.append(colspec)
+            colspecs.append(spec_from_field(field))
 
         if gt_field is not None:
             ploidy = gt_field.summary.max_number - 1
             shape = [m, n]
             chunks = [variants_chunk_size, samples_chunk_size]
             dimensions = ["variants", "samples"]
-
             colspecs.append(
                 ZarrColumnSpec.new(
                     vcf_field=None,
@@ -1457,12 +1527,11 @@ class VcfZarrSchema:
             format_version=ZARR_SCHEMA_FORMAT_VERSION,
             samples_chunk_size=samples_chunk_size,
             variants_chunk_size=variants_chunk_size,
-            columns={col.name: col for col in colspecs},
+            fields={col.name: col for col in colspecs},
             dimensions=["variants", "samples", "ploidy", "alleles", "filters"],
-            sample_id=icf.metadata.samples,
-            contig_id=icf.metadata.contig_names,
-            contig_length=icf.metadata.contig_lengths,
-            filter_id=icf.metadata.filters,
+            samples=icf.metadata.samples,
+            contigs=icf.metadata.contigs,
+            filters=icf.metadata.filters,
         )
 
 
@@ -1470,14 +1539,12 @@ class VcfZarr:
     def __init__(self, path):
         if not (path / ".zmetadata").exists():
             raise ValueError("Not in VcfZarr format")  # NEEDS TEST
+        self.path = path
         self.root = zarr.open(path, mode="r")
 
-    def __repr__(self):
-        return repr(self.root)  # NEEDS TEST
-
     def summary_table(self):
         data = []
-        arrays = [(a.nbytes_stored, a) for _, a in self.root.arrays()]
+        arrays = [(core.du(self.path / a.basename), a) for _, a in self.root.arrays()]
         arrays.sort(key=lambda x: x[0])
         for stored, array in reversed(arrays):
             d = {
@@ -1498,15 +1565,6 @@ class VcfZarr:
         return data
 
 
-@dataclasses.dataclass
-class EncodingWork:
-    func: callable = dataclasses.field(repr=False)
-    start: int
-    stop: int
-    columns: list[str]
-    memory: int = 0
-
-
 def parse_max_memory(max_memory):
     if max_memory is None:
         # Effectively unbounded
@@ -1517,67 +1575,299 @@ def parse_max_memory(max_memory):
     return max_memory
 
 
+@dataclasses.dataclass
+class VcfZarrPartition:
+    start: int
+    stop: int
+
+    @staticmethod
+    def generate_partitions(num_records, chunk_size, num_partitions, max_chunks=None):
+        num_chunks = int(np.ceil(num_records / chunk_size))
+        if max_chunks is not None:
+            num_chunks = min(num_chunks, max_chunks)
+        partitions = []
+        splits = np.array_split(np.arange(num_chunks), min(num_partitions, num_chunks))
+        for chunk_slice in splits:
+            start_chunk = int(chunk_slice[0])
+            stop_chunk = int(chunk_slice[-1]) + 1
+            start_index = start_chunk * chunk_size
+            stop_index = min(stop_chunk * chunk_size, num_records)
+            partitions.append(VcfZarrPartition(start_index, stop_index))
+        return partitions
+
+
+VZW_METADATA_FORMAT_VERSION = "0.1"
+
+
+@dataclasses.dataclass
+class VcfZarrWriterMetadata:
+    format_version: str
+    icf_path: str
+    schema: VcfZarrSchema
+    dimension_separator: str
+    partitions: list
+    provenance: dict
+
+    def asdict(self):
+        return dataclasses.asdict(self)
+
+    @staticmethod
+    def fromdict(d):
+        if d["format_version"] != VZW_METADATA_FORMAT_VERSION:
+            raise ValueError(
+                "VcfZarrWriter format version mismatch: "
+                f"{d['format_version']} != {VZW_METADATA_FORMAT_VERSION}"
+            )
+        ret = VcfZarrWriterMetadata(**d)
+        ret.schema = VcfZarrSchema.fromdict(ret.schema)
+        ret.partitions = [VcfZarrPartition(**p) for p in ret.partitions]
+        return ret
+
+
 class VcfZarrWriter:
-    def __init__(self, path, icf, schema, dimension_separator=None):
+    def __init__(self, path):
         self.path = pathlib.Path(path)
+        self.wip_path = self.path / "wip"
+        self.arrays_path = self.wip_path / "arrays"
+        self.partitions_path = self.wip_path / "partitions"
+        self.metadata = None
+        self.icf = None
+
+    @property
+    def schema(self):
+        return self.metadata.schema
+
+    @property
+    def num_partitions(self):
+        return len(self.metadata.partitions)
+
+    #######################
+    # init
+    #######################
+
+    def init(
+        self,
+        icf,
+        *,
+        target_num_partitions,
+        schema,
+        dimension_separator=None,
+        max_variant_chunks=None,
+    ):
         self.icf = icf
-        self.schema = schema
+        if self.path.exists():
+            raise ValueError("Zarr path already exists")  # NEEDS TEST
+        partitions = VcfZarrPartition.generate_partitions(
+            self.icf.num_records,
+            schema.variants_chunk_size,
+            target_num_partitions,
+            max_chunks=max_variant_chunks,
+        )
         # Default to using nested directories following the Zarr v3 default.
         # This seems to require version 2.17+ to work properly
-        self.dimension_separator = (
+        dimension_separator = (
             "/" if dimension_separator is None else dimension_separator
         )
+        self.metadata = VcfZarrWriterMetadata(
+            format_version=VZW_METADATA_FORMAT_VERSION,
+            icf_path=str(self.icf.path),
+            schema=schema,
+            dimension_separator=dimension_separator,
+            partitions=partitions,
+            # Bare minimum here for provenance - see comments above
+            provenance={"source": f"bio2zarr-{provenance.__version__}"},
+        )
+
+        self.path.mkdir()
         store = zarr.DirectoryStore(self.path)
-        self.root = zarr.group(store=store)
+        root = zarr.group(store=store)
+        root.attrs.update(
+            {
+                "vcf_zarr_version": "0.2",
+                "vcf_header": self.icf.vcf_header,
+                "source": f"bio2zarr-{provenance.__version__}",
+            }
+        )
+        # Doing this syncronously - this is fine surely
+        self.encode_samples(root)
+        self.encode_filter_id(root)
+        self.encode_contig_id(root)
 
-    def init_array(self, variable):
+        self.wip_path.mkdir()
+        self.arrays_path.mkdir()
+        self.partitions_path.mkdir()
+        store = zarr.DirectoryStore(self.arrays_path)
+        root = zarr.group(store=store)
+
+        for column in self.schema.fields.values():
+            self.init_array(root, column, partitions[-1].stop)
+
+        logger.info("Writing WIP metadata")
+        with open(self.wip_path / "metadata.json", "w") as f:
+            json.dump(self.metadata.asdict(), f, indent=4)
+        return len(partitions)
+
+    def encode_samples(self, root):
+        if self.schema.samples != self.icf.metadata.samples:
+            raise ValueError(
+                "Subsetting or reordering samples not supported currently"
+            )  # NEEDS TEST
+        array = root.array(
+            "sample_id",
+            [sample.id for sample in self.schema.samples],
+            dtype="str",
+            compressor=DEFAULT_ZARR_COMPRESSOR,
+            chunks=(self.schema.samples_chunk_size,),
+        )
+        array.attrs["_ARRAY_DIMENSIONS"] = ["samples"]
+        logger.debug("Samples done")
+
+    def encode_contig_id(self, root):
+        array = root.array(
+            "contig_id",
+            [contig.id for contig in self.schema.contigs],
+            dtype="str",
+            compressor=DEFAULT_ZARR_COMPRESSOR,
+        )
+        array.attrs["_ARRAY_DIMENSIONS"] = ["contigs"]
+        if all(contig.length is not None for contig in self.schema.contigs):
+            array = root.array(
+                "contig_length",
+                [contig.length for contig in self.schema.contigs],
+                dtype=np.int64,
+                compressor=DEFAULT_ZARR_COMPRESSOR,
+            )
+            array.attrs["_ARRAY_DIMENSIONS"] = ["contigs"]
+
+    def encode_filter_id(self, root):
+        # TODO need a way to store description also
+        # https://github.com/sgkit-dev/vcf-zarr-spec/issues/19
+        array = root.array(
+            "filter_id",
+            [filt.id for filt in self.schema.filters],
+            dtype="str",
+            compressor=DEFAULT_ZARR_COMPRESSOR,
+        )
+        array.attrs["_ARRAY_DIMENSIONS"] = ["filters"]
+
+    def init_array(self, root, variable, variants_dim_size):
         object_codec = None
         if variable.dtype == "O":
             object_codec = numcodecs.VLenUTF8()
-        a = self.root.empty(
-            "wip_" + variable.name,
-            shape=variable.shape,
+        shape = list(variable.shape)
+        # Truncate the variants dimension is max_variant_chunks was specified
+        shape[0] = variants_dim_size
+        a = root.empty(
+            variable.name,
+            shape=shape,
             chunks=variable.chunks,
             dtype=variable.dtype,
             compressor=numcodecs.get_codec(variable.compressor),
             filters=[numcodecs.get_codec(filt) for filt in variable.filters],
             object_codec=object_codec,
-            dimension_separator=self.dimension_separator,
+            dimension_separator=self.metadata.dimension_separator,
         )
-        # Dimension names are part of the spec in Zarr v3
-        a.attrs["_ARRAY_DIMENSIONS"] = variable.dimensions
-
-    def get_array(self, name):
-        return self.root["wip_" + name]
-
-    def finalise_array(self, variable_name):
-        source = self.path / ("wip_" + variable_name)
-        dest = self.path / variable_name
-        # Atomic swap
-        os.rename(source, dest)
-        logger.info(f"Finalised {variable_name}")
-
-    def encode_array_slice(self, column, start, stop):
-        source_col = self.icf.columns[column.vcf_field]
-        array = self.get_array(column.name)
-        ba = core.BufferedArray(array, start)
+        a.attrs.update(
+            {
+                "description": variable.description,
+                # Dimension names are part of the spec in Zarr v3
+                "_ARRAY_DIMENSIONS": variable.dimensions,
+            }
+        )
+        logger.debug(f"Initialised {a}")
+
+    #######################
+    # encode_partition
+    #######################
+
+    def load_metadata(self):
+        if self.metadata is None:
+            with open(self.wip_path / "metadata.json") as f:
+                self.metadata = VcfZarrWriterMetadata.fromdict(json.load(f))
+            self.icf = IntermediateColumnarFormat(self.metadata.icf_path)
+
+    def partition_path(self, partition_index):
+        return self.partitions_path / f"p{partition_index}"
+
+    def wip_partition_path(self, partition_index):
+        return self.partitions_path / f"wip_p{partition_index}"
+
+    def wip_partition_array_path(self, partition_index, name):
+        return self.wip_partition_path(partition_index) / name
+
+    def partition_array_path(self, partition_index, name):
+        return self.partition_path(partition_index) / name
+
+    def encode_partition(self, partition_index):
+        self.load_metadata()
+        if partition_index < 0 or partition_index >= self.num_partitions:
+            raise ValueError(
+                "Partition index must be in the range 0 <= index < num_partitions"
+            )
+        partition_path = self.wip_partition_path(partition_index)
+        partition_path.mkdir(exist_ok=True)
+        logger.info(f"Encoding partition {partition_index} to {partition_path}")
+
+        self.encode_id_partition(partition_index)
+        self.encode_filters_partition(partition_index)
+        self.encode_contig_partition(partition_index)
+        self.encode_alleles_partition(partition_index)
+        for col in self.schema.fields.values():
+            if col.vcf_field is not None:
+                self.encode_array_partition(col, partition_index)
+        if "call_genotype" in self.schema.fields:
+            self.encode_genotypes_partition(partition_index)
+
+        final_path = self.partition_path(partition_index)
+        logger.info(f"Finalising {partition_index} at {final_path}")
+        if final_path.exists():
+            logger.warning(f"Removing existing partition at {final_path}")
+            shutil.rmtree(final_path)
+        os.rename(partition_path, final_path)
+
+    def init_partition_array(self, partition_index, name):
+        wip_path = self.wip_partition_array_path(partition_index, name)
+        # Create an empty array like the definition
+        src = self.arrays_path / name
+        # Overwrite any existing WIP files
+        shutil.copytree(src, wip_path, dirs_exist_ok=True)
+        array = zarr.open(wip_path)
+        logger.debug(f"Opened empty array {array} @ {wip_path}")
+        return array
+
+    def finalise_partition_array(self, partition_index, name):
+        logger.debug(f"Encoded {name} partition {partition_index}")
+
+    def encode_array_partition(self, column, partition_index):
+        array = self.init_partition_array(partition_index, column.name)
+
+        partition = self.metadata.partitions[partition_index]
+        ba = core.BufferedArray(array, partition.start)
+        source_col = self.icf.fields[column.vcf_field]
         sanitiser = source_col.sanitiser_factory(ba.buff.shape)
 
-        for value in source_col.iter_values(start, stop):
+        for value in source_col.iter_values(partition.start, partition.stop):
             # We write directly into the buffer in the sanitiser function
             # to make it easier to reason about dimension padding
             j = ba.next_buffer_row()
             sanitiser(ba.buff, j, value)
         ba.flush()
-        logger.debug(f"Encoded {column.name} slice {start}:{stop}")
+        self.finalise_partition_array(partition_index, column.name)
+
+    def encode_genotypes_partition(self, partition_index):
+        gt_array = self.init_partition_array(partition_index, "call_genotype")
+        gt_mask_array = self.init_partition_array(partition_index, "call_genotype_mask")
+        gt_phased_array = self.init_partition_array(
+            partition_index, "call_genotype_phased"
+        )
 
-    def encode_genotypes_slice(self, start, stop):
-        source_col = self.icf.columns["FORMAT/GT"]
-        gt = core.BufferedArray(self.get_array("call_genotype"), start)
-        gt_mask = core.BufferedArray(self.get_array("call_genotype_mask"), start)
-        gt_phased = core.BufferedArray(self.get_array("call_genotype_phased"), start)
+        partition = self.metadata.partitions[partition_index]
+        gt = core.BufferedArray(gt_array, partition.start)
+        gt_mask = core.BufferedArray(gt_mask_array, partition.start)
+        gt_phased = core.BufferedArray(gt_phased_array, partition.start)
 
-        for value in source_col.iter_values(start, stop):
+        source_col = self.icf.fields["FORMAT/GT"]
+        for value in source_col.iter_values(partition.start, partition.stop):
             j = gt.next_buffer_row()
             sanitise_value_int_2d(gt.buff, j, value[:, :-1])
             j = gt_phased.next_buffer_row()
@@ -1589,29 +1879,40 @@ class VcfZarrWriter:
         gt.flush()
         gt_phased.flush()
         gt_mask.flush()
-        logger.debug(f"Encoded GT slice {start}:{stop}")
 
-    def encode_alleles_slice(self, start, stop):
-        ref_col = self.icf.columns["REF"]
-        alt_col = self.icf.columns["ALT"]
-        alleles = core.BufferedArray(self.get_array("variant_allele"), start)
+        self.finalise_partition_array(partition_index, "call_genotype")
+        self.finalise_partition_array(partition_index, "call_genotype_mask")
+        self.finalise_partition_array(partition_index, "call_genotype_phased")
+
+    def encode_alleles_partition(self, partition_index):
+        array_name = "variant_allele"
+        alleles_array = self.init_partition_array(partition_index, array_name)
+        partition = self.metadata.partitions[partition_index]
+        alleles = core.BufferedArray(alleles_array, partition.start)
+        ref_col = self.icf.fields["REF"]
+        alt_col = self.icf.fields["ALT"]
 
         for ref, alt in zip(
-            ref_col.iter_values(start, stop), alt_col.iter_values(start, stop)
+            ref_col.iter_values(partition.start, partition.stop),
+            alt_col.iter_values(partition.start, partition.stop),
         ):
             j = alleles.next_buffer_row()
             alleles.buff[j, :] = STR_FILL
             alleles.buff[j, 0] = ref[0]
             alleles.buff[j, 1 : 1 + len(alt)] = alt
         alleles.flush()
-        logger.debug(f"Encoded alleles slice {start}:{stop}")
 
-    def encode_id_slice(self, start, stop):
-        col = self.icf.columns["ID"]
-        vid = core.BufferedArray(self.get_array("variant_id"), start)
-        vid_mask = core.BufferedArray(self.get_array("variant_id_mask"), start)
+        self.finalise_partition_array(partition_index, array_name)
 
-        for value in col.iter_values(start, stop):
+    def encode_id_partition(self, partition_index):
+        vid_array = self.init_partition_array(partition_index, "variant_id")
+        vid_mask_array = self.init_partition_array(partition_index, "variant_id_mask")
+        partition = self.metadata.partitions[partition_index]
+        vid = core.BufferedArray(vid_array, partition.start)
+        vid_mask = core.BufferedArray(vid_mask_array, partition.start)
+        col = self.icf.fields["ID"]
+
+        for value in col.iter_values(partition.start, partition.stop):
             j = vid.next_buffer_row()
             k = vid_mask.next_buffer_row()
             assert j == k
@@ -1623,13 +1924,19 @@ class VcfZarrWriter:
                 vid_mask.buff[j] = True
         vid.flush()
         vid_mask.flush()
-        logger.debug(f"Encoded ID slice {start}:{stop}")
 
-    def encode_filters_slice(self, lookup, start, stop):
-        col = self.icf.columns["FILTERS"]
-        var_filter = core.BufferedArray(self.get_array("variant_filter"), start)
+        self.finalise_partition_array(partition_index, "variant_id")
+        self.finalise_partition_array(partition_index, "variant_id_mask")
+
+    def encode_filters_partition(self, partition_index):
+        lookup = {filt.id: index for index, filt in enumerate(self.schema.filters)}
+        array_name = "variant_filter"
+        array = self.init_partition_array(partition_index, array_name)
+        partition = self.metadata.partitions[partition_index]
+        var_filter = core.BufferedArray(array, partition.start)
 
-        for value in col.iter_values(start, stop):
+        col = self.icf.fields["FILTERS"]
+        for value in col.iter_values(partition.start, partition.stop):
             j = var_filter.next_buffer_row()
             var_filter.buff[j] = False
             for f in value:
@@ -1637,16 +1944,21 @@ class VcfZarrWriter:
                     var_filter.buff[j, lookup[f]] = True
                 except KeyError:
                     raise ValueError(
-                        f"Filter '{f}' was not defined " f"in the header."
+                        f"Filter '{f}' was not defined in the header."
                     ) from None
         var_filter.flush()
-        logger.debug(f"Encoded FILTERS slice {start}:{stop}")
 
-    def encode_contig_slice(self, lookup, start, stop):
-        col = self.icf.columns["CHROM"]
-        contig = core.BufferedArray(self.get_array("variant_contig"), start)
+        self.finalise_partition_array(partition_index, array_name)
+
+    def encode_contig_partition(self, partition_index):
+        lookup = {contig.id: index for index, contig in enumerate(self.schema.contigs)}
+        array_name = "variant_contig"
+        array = self.init_partition_array(partition_index, array_name)
+        partition = self.metadata.partitions[partition_index]
+        contig = core.BufferedArray(array, partition.start)
+        col = self.icf.fields["CHROM"]
 
-        for value in col.iter_values(start, stop):
+        for value in col.iter_values(partition.start, partition.stop):
             j = contig.next_buffer_row()
             # Note: because we are using the indexes to define the lookups
             # and we always have an index, it seems that we the contig lookup
@@ -1654,161 +1966,131 @@ class VcfZarrWriter:
             # here, please do open an issue with a reproducible example!
             contig.buff[j] = lookup[value[0]]
         contig.flush()
-        logger.debug(f"Encoded CHROM slice {start}:{stop}")
 
-    def encode_samples(self):
-        if not np.array_equal(self.schema.sample_id, self.icf.metadata.samples):
-            raise ValueError(
-                "Subsetting or reordering samples not supported currently"
-            )  # NEEDS TEST
-        array = self.root.array(
-            "sample_id",
-            self.schema.sample_id,
-            dtype="str",
-            compressor=DEFAULT_ZARR_COMPRESSOR,
-            chunks=(self.schema.samples_chunk_size,),
-        )
-        array.attrs["_ARRAY_DIMENSIONS"] = ["samples"]
-        logger.debug("Samples done")
-
-    def encode_contig_id(self):
-        array = self.root.array(
-            "contig_id",
-            self.schema.contig_id,
-            dtype="str",
-            compressor=DEFAULT_ZARR_COMPRESSOR,
-        )
-        array.attrs["_ARRAY_DIMENSIONS"] = ["contigs"]
-        if self.schema.contig_length is not None:
-            array = self.root.array(
-                "contig_length",
-                self.schema.contig_length,
-                dtype=np.int64,
-                compressor=DEFAULT_ZARR_COMPRESSOR,
+        self.finalise_partition_array(partition_index, array_name)
+
+    #######################
+    # finalise
+    #######################
+
+    def finalise_array(self, name):
+        logger.info(f"Finalising {name}")
+        final_path = self.path / name
+        if final_path.exists():
+            # NEEDS TEST
+            raise ValueError(f"Array {name} already exists")
+        for partition in range(self.num_partitions):
+            # Move all the files in partition dir to dest dir
+            src = self.partition_array_path(partition, name)
+            if not src.exists():
+                # Needs test
+                raise ValueError(f"Partition {partition} of {name} does not exist")
+            dest = self.arrays_path / name
+            # This is Zarr v2 specific. Chunks in v3 with start with "c" prefix.
+            chunk_files = [
+                path for path in src.iterdir() if not path.name.startswith(".")
+            ]
+            # TODO check for a count of then number of files. If we require a
+            # dimension_separator of "/" then we could make stronger assertions
+            # here, as we'd always have num_variant_chunks
+            logger.debug(
+                f"Moving {len(chunk_files)} chunks for {name} partition {partition}"
             )
-            array.attrs["_ARRAY_DIMENSIONS"] = ["contigs"]
-        return {v: j for j, v in enumerate(self.schema.contig_id)}
+            for chunk_file in chunk_files:
+                os.rename(chunk_file, dest / chunk_file.name)
+        # Finally, once all the chunks have moved into the arrays dir,
+        # we move it out of wip
+        os.rename(self.arrays_path / name, self.path / name)
+        core.update_progress(1)
 
-    def encode_filter_id(self):
-        array = self.root.array(
-            "filter_id",
-            self.schema.filter_id,
-            dtype="str",
-            compressor=DEFAULT_ZARR_COMPRESSOR,
-        )
-        array.attrs["_ARRAY_DIMENSIONS"] = ["filters"]
-        return {v: j for j, v in enumerate(self.schema.filter_id)}
+    def finalise(self, show_progress=False):
+        self.load_metadata()
 
-    def init(self):
-        self.root.attrs["vcf_zarr_version"] = "0.2"
-        self.root.attrs["vcf_header"] = self.icf.vcf_header
-        self.root.attrs["source"] = f"bio2zarr-{provenance.__version__}"
-        for column in self.schema.columns.values():
-            self.init_array(column)
+        logger.info("Scanning {self.num_partitions} partitions")
+        missing = []
+        # TODO may need a progress bar here
+        for partition_id in range(self.num_partitions):
+            if not self.partition_path(partition_id).exists():
+                missing.append(partition_id)
+        if len(missing) > 0:
+            raise FileNotFoundError(f"Partitions not encoded: {missing}")
 
-    def finalise(self):
+        progress_config = core.ProgressConfig(
+            total=len(self.schema.fields),
+            title="Finalise",
+            units="array",
+            show=show_progress,
+        )
+        # NOTE: it's not clear that adding more workers will make this quicker,
+        # as it's just going to be causing contention on the file system.
+        # Something to check empirically in some deployments.
+        # FIXME we're just using worker_processes=0 here to hook into the
+        # SynchronousExecutor which is intended for testing purposes so
+        # that we get test coverage. Should fix this either by allowing
+        # for multiple workers, or making a standard wrapper for tqdm
+        # that allows us to have a consistent look and feel.
+        with core.ParallelWorkManager(0, progress_config) as pwm:
+            for name in self.schema.fields:
+                pwm.submit(self.finalise_array, name)
+        logger.debug(f"Removing {self.wip_path}")
+        shutil.rmtree(self.wip_path)
+        logger.info("Consolidating Zarr metadata")
         zarr.consolidate_metadata(self.path)
 
-    def encode(
-        self,
-        worker_processes=1,
-        max_v_chunks=None,
-        show_progress=False,
-        max_memory=None,
-    ):
-        max_memory = parse_max_memory(max_memory)
+    ######################
+    # encode_all_partitions
+    ######################
 
-        # TODO this will move into the setup logic later when we're making it possible
-        # to split the work by slice
-        num_slices = max(1, worker_processes * 4)
-        # Using POS arbitrarily to get the array slices
-        slices = core.chunk_aligned_slices(
-            self.get_array("variant_position"), num_slices, max_chunks=max_v_chunks
+    def get_max_encoding_memory(self):
+        """
+        Return the approximate maximum memory used to encode a variant chunk.
+        """
+        max_encoding_mem = max(
+            col.variant_chunk_nbytes for col in self.schema.fields.values()
         )
-        truncated = slices[-1][-1]
-        for array in self.root.values():
-            if array.attrs["_ARRAY_DIMENSIONS"][0] == "variants":
-                shape = list(array.shape)
-                shape[0] = truncated
-                array.resize(shape)
-
-        total_bytes = 0
-        encoding_memory_requirements = {}
-        for col in self.schema.columns.values():
-            array = self.get_array(col.name)
-            # NOTE!! this is bad, we're potentially creating quite a large
-            # numpy array for basically nothing. We can compute this.
-            variant_chunk_size = array.blocks[0].nbytes
-            encoding_memory_requirements[col.name] = variant_chunk_size
-            logger.debug(
-                f"{col.name} requires at least {display_size(variant_chunk_size)} "
-                f"per worker"
+        gt_mem = 0
+        if "call_genotype" in self.schema.fields:
+            encoded_together = [
+                "call_genotype",
+                "call_genotype_phased",
+                "call_genotype_mask",
+            ]
+            gt_mem = sum(
+                self.schema.fields[col].variant_chunk_nbytes for col in encoded_together
             )
-            total_bytes += array.nbytes
-
-        filter_id_map = self.encode_filter_id()
-        contig_id_map = self.encode_contig_id()
-
-        work = []
-        for start, stop in slices:
-            for col in self.schema.columns.values():
-                if col.vcf_field is not None:
-                    f = functools.partial(self.encode_array_slice, col)
-                    work.append(
-                        EncodingWork(
-                            f,
-                            start,
-                            stop,
-                            [col.name],
-                            encoding_memory_requirements[col.name],
-                        )
-                    )
-            work.append(
-                EncodingWork(self.encode_alleles_slice, start, stop, ["variant_allele"])
-            )
-            work.append(
-                EncodingWork(
-                    self.encode_id_slice, start, stop, ["variant_id", "variant_id_mask"]
-                )
-            )
-            work.append(
-                EncodingWork(
-                    functools.partial(self.encode_filters_slice, filter_id_map),
-                    start,
-                    stop,
-                    ["variant_filter"],
-                )
+        return max(max_encoding_mem, gt_mem)
+
+    def encode_all_partitions(
+        self, *, worker_processes=1, show_progress=False, max_memory=None
+    ):
+        max_memory = parse_max_memory(max_memory)
+        self.load_metadata()
+        num_partitions = self.num_partitions
+        per_worker_memory = self.get_max_encoding_memory()
+        logger.info(
+            f"Encoding Zarr over {num_partitions} partitions with "
+            f"{worker_processes} workers and {display_size(per_worker_memory)} "
+            "per worker"
+        )
+        # Each partition requires per_worker_memory bytes, so to prevent more that
+        # max_memory being used, we clamp the number of workers
+        max_num_workers = max_memory // per_worker_memory
+        if max_num_workers < worker_processes:
+            logger.warning(
+                f"Limiting number of workers to {max_num_workers} to "
+                f"keep within specified memory budget of {display_size(max_memory)}"
             )
-            work.append(
-                EncodingWork(
-                    functools.partial(self.encode_contig_slice, contig_id_map),
-                    start,
-                    stop,
-                    ["variant_contig"],
-                )
+        if max_num_workers <= 0:
+            raise ValueError(
+                f"Insufficient memory to encode a partition:"
+                f"{display_size(per_worker_memory)} > {display_size(max_memory)}"
             )
-            if "call_genotype" in self.schema.columns:
-                variables = [
-                    "call_genotype",
-                    "call_genotype_phased",
-                    "call_genotype_mask",
-                ]
-                gt_memory = sum(
-                    encoding_memory_requirements[name] for name in variables
-                )
-                work.append(
-                    EncodingWork(
-                        self.encode_genotypes_slice, start, stop, variables, gt_memory
-                    )
-                )
+        num_workers = min(max_num_workers, worker_processes)
 
-        # Fail early if we can't fit a particular column into memory
-        for wp in work:
-            if wp.memory > max_memory:
-                raise ValueError(
-                    f"Insufficient memory for {wp.columns}: "
-                    f"{display_size(wp.memory)} > {display_size(max_memory)}"
-                )
+        total_bytes = 0
+        for col in self.schema.fields.values():
+            # Open the array definition to get the total size
+            total_bytes += zarr.open(self.arrays_path / col.name).nbytes
 
         progress_config = core.ProgressConfig(
             total=total_bytes,
@@ -1816,54 +2098,9 @@ class VcfZarrWriter:
             units="B",
             show=show_progress,
         )
-
-        used_memory = 0
-        # We need to keep some bounds on the queue size or the memory bounds algorithm
-        # below doesn't really work.
-        max_queued = 4 * max(1, worker_processes)
-        encoded_slices = collections.Counter()
-
-        with core.ParallelWorkManager(worker_processes, progress_config) as pwm:
-            future = pwm.submit(self.encode_samples)
-            future_to_work = {future: EncodingWork(None, 0, 0, [])}
-
-            def service_completed_futures():
-                nonlocal used_memory
-
-                completed = pwm.wait_for_completed()
-                for future in completed:
-                    wp_done = future_to_work.pop(future)
-                    used_memory -= wp_done.memory
-                    logger.debug(
-                        f"Complete {wp_done}: used mem={display_size(used_memory)}"
-                    )
-                    for column in wp_done.columns:
-                        encoded_slices[column] += 1
-                        if encoded_slices[column] == len(slices):
-                            # Do this syncronously for simplicity. Should be
-                            # fine as the workers will probably be busy with
-                            # large encode tasks most of the time.
-                            self.finalise_array(column)
-
-            for wp in work:
-                while (
-                    used_memory + wp.memory > max_memory
-                    or len(future_to_work) > max_queued
-                ):
-                    logger.debug(
-                        f"Wait: mem_required={used_memory + wp.memory} "
-                        f"max_mem={max_memory} queued={len(future_to_work)} "
-                        f"max_queued={max_queued}"
-                    )
-                    service_completed_futures()
-                future = pwm.submit(wp.func, wp.start, wp.stop)
-                used_memory += wp.memory
-                logger.debug(f"Submit {wp}: used mem={display_size(used_memory)}")
-                future_to_work[future] = wp
-
-            logger.debug("All work submitted")
-            while len(future_to_work) > 0:
-                service_completed_futures()
+        with core.ParallelWorkManager(num_workers, progress_config) as pwm:
+            for partition_index in range(num_partitions):
+                pwm.submit(self.encode_partition, partition_index)
 
 
 def mkschema(if_path, out):
@@ -1878,13 +2115,48 @@ def encode(
     schema_path=None,
     variants_chunk_size=None,
     samples_chunk_size=None,
-    max_v_chunks=None,
+    max_variant_chunks=None,
     dimension_separator=None,
     max_memory=None,
     worker_processes=1,
     show_progress=False,
 ):
-    icf = IntermediateColumnarFormat(if_path)
+    # Rough heuristic to split work up enough to keep utilisation high
+    target_num_partitions = max(1, worker_processes * 4)
+    encode_init(
+        if_path,
+        zarr_path,
+        target_num_partitions,
+        schema_path=schema_path,
+        variants_chunk_size=variants_chunk_size,
+        samples_chunk_size=samples_chunk_size,
+        max_variant_chunks=max_variant_chunks,
+        dimension_separator=dimension_separator,
+    )
+    vzw = VcfZarrWriter(zarr_path)
+    vzw.encode_all_partitions(
+        worker_processes=worker_processes,
+        show_progress=show_progress,
+        max_memory=max_memory,
+    )
+    vzw.finalise(show_progress)
+
+
+def encode_init(
+    icf_path,
+    zarr_path,
+    target_num_partitions,
+    *,
+    schema_path=None,
+    variants_chunk_size=None,
+    samples_chunk_size=None,
+    max_variant_chunks=None,
+    dimension_separator=None,
+    max_memory=None,
+    worker_processes=1,
+    show_progress=False,
+):
+    icf = IntermediateColumnarFormat(icf_path)
     if schema_path is None:
         schema = VcfZarrSchema.generate(
             icf,
@@ -1900,18 +2172,25 @@ def encode(
         with open(schema_path) as f:
             schema = VcfZarrSchema.fromjson(f.read())
     zarr_path = pathlib.Path(zarr_path)
-    if zarr_path.exists():
-        logger.warning(f"Deleting existing {zarr_path}")
-        shutil.rmtree(zarr_path)
-    vzw = VcfZarrWriter(zarr_path, icf, schema, dimension_separator=dimension_separator)
-    vzw.init()
-    vzw.encode(
-        max_v_chunks=max_v_chunks,
-        worker_processes=worker_processes,
-        max_memory=max_memory,
-        show_progress=show_progress,
+    vzw = VcfZarrWriter(zarr_path)
+    vzw.init(
+        icf,
+        target_num_partitions=target_num_partitions,
+        schema=schema,
+        dimension_separator=dimension_separator,
+        max_variant_chunks=max_variant_chunks,
     )
-    vzw.finalise()
+    return vzw.num_partitions, vzw.get_max_encoding_memory()
+
+
+def encode_partition(zarr_path, partition):
+    writer = VcfZarrWriter(zarr_path)
+    writer.encode_partition(partition)
+
+
+def encode_finalise(zarr_path, show_progress=False):
+    writer = VcfZarrWriter(zarr_path)
+    writer.finalise(show_progress=show_progress)
 
 
 def convert(
@@ -2121,7 +2400,7 @@ def validate(vcf_path, zarr_path, show_progress=False):
     assert pos[start_index] == first_pos
     vcf = cyvcf2.VCF(vcf_path)
     if show_progress:
-        iterator = tqdm.tqdm(vcf, desc=" Verify", total=vcf.num_records)  # NEEDS TEST
+        iterator = tqdm.tqdm(vcf, desc="  Verify", total=vcf.num_records)  # NEEDS TEST
     else:
         iterator = vcf
     for j, row in enumerate(iterator, start_index):