diff --git a/.gitignore b/.gitignore
index 6786d43..86578cd 100644
--- a/.gitignore
+++ b/.gitignore
@@ -7,4 +7,5 @@ Manifest*.toml
 ref/
 .claude/
 data/.*
-data/*
\ No newline at end of file
+data/*
+!/data/a_sparse_cdf.cdf
diff --git a/data/a_sparse_cdf.cdf b/data/a_sparse_cdf.cdf
new file mode 100644
index 0000000..89814fb
Binary files /dev/null and b/data/a_sparse_cdf.cdf differ
diff --git a/src/loading/variable.jl b/src/loading/variable.jl
index 020042b..93b07e6 100644
--- a/src/loading/variable.jl
+++ b/src/loading/variable.jl
@@ -116,12 +116,11 @@ function DiskArrays.readblock!(var::CDFVariable{T, N}, dest::AbstractArray{T}, r
     buffer = parent(var.parentdataset)
     RecordSizeType = recordsize_type(var.parentdataset)
     entries, vvr_type = read_vvrs(var.vdr)
+    # vvr record type is the ultimate source of compression
+    compression = vvr_type == VVR_ ? NoCompression : variable_compression(var.vdr)
+    sparse_type(var.vdr) == 0 ||
+        return _readblock_sparse!(var, dest, ranges, entries, compression, buffer, RecordSizeType)
     isempty(entries) && return dest
-    compression = if !isempty(entries) #  # vvr records is the ultimative source
-        vvr_type == VVR_ ? NoCompression : variable_compression(var.vdr)
-    else
-        NoCompression
-    end
 
     record_range = ranges[end]
     other_ranges = ranges[1:(N - 1)]
@@ -195,6 +194,65 @@ function DiskArrays.readblock!(var::CDFVariable{T, N}, dest::AbstractArray{T}, r
     return dest
 end
 
+# Sparse records: records absent from the VXR are virtual. Pad sparse (1) fills them
+# with the VDR pad value (or the NASA default pad); previous sparse (2) repeats the
+# last record of the preceding physical block.
+# Mirrors cdflib semantics; spec'd in the CDF IFD (sRecords).
+function _readblock_sparse!(var::CDFVariable{T, N}, dest, ranges, entries, compression, buffer, ::Type{RST}) where {T, N, RST}
+    record_range = ranges[end]
+    other_ranges = ranges[1:(N - 1)]
+    dims_without_record = var.dims[1:(N - 1)]
+    record_size = prod(dims_without_record)
+    is_row_major = majority(var) == Row
+    needs_byte_swap = is_big_endian_encoding(var)
+    use_prev = sparse_type(var.vdr) == 2
+    pad = pad_value(var.vdr, T, needs_byte_swap)
+
+    chunk = Vector{T}()
+    cached = 0
+    # load (and permute, for row-major) a physical block once; runs of records from
+    # the same block (incl. prev-sparse repeats) reuse it
+    load_block = function (idx)
+        entry = entries[idx]
+        if cached != idx
+            resize!(chunk, record_size * length(entry))
+            if compression == NoCompression
+                load_vvr_data!(chunk, 1, buffer, entry.offset, length(chunk), RST)
+            else
+                decompressor = take!(decompressors())
+                try
+                    load_cvvr_data!(chunk, 1, buffer, entry.offset, length(chunk), RST, compression; decompressor)
+                finally
+                    put!(decompressors(), decompressor)
+                end
+            end
+            is_row_major && majority_swap!(reshape(chunk, dims_without_record..., :), dims_without_record)
+            cached = idx
+        end
+        return reshape(chunk, dims_without_record..., :)
+    end
+
+    blk = 1
+    nblocks = length(entries)
+    for (di, r) in enumerate(record_range)
+        while blk <= nblocks && entries[blk].last < r
+            blk += 1
+        end
+        dest_view = _record_view(dest, di)
+        if blk <= nblocks && entries[blk].first <= r
+            arr = load_block(blk)
+            dest_view .= view(arr, other_ranges..., r - entries[blk].first + 1)
+        elseif use_prev && blk > 1
+            arr = load_block(blk - 1)
+            dest_view .= view(arr, other_ranges..., size(arr, N))
+        else
+            fill!(dest_view, pad)
+        end
+    end
+    needs_byte_swap && _byte_swap!(dest)
+    return dest
+end
+
 function collect_vxr_entries!(entries::Vector{VVREntry}, src, offset, ::Type{FieldSizeT}) where {FieldSizeT}
     vvr_type = nothing
     while offset != 0
diff --git a/src/records/vdr.jl b/src/records/vdr.jl
index 47b1eae..8a88562 100644
--- a/src/records/vdr.jl
+++ b/src/records/vdr.jl
@@ -154,4 +154,40 @@ end
 is_record_varying(vdr) = !is_nrv(vdr)
 """Whether or not the variable is a non-record variable"""
 is_nrv(vdr) = (vdr.flags & 0x01) == 0
+has_pad_value(vdr::AbstractVDR) = (vdr.flags & 0x02) != 0
 is_compressed(vdr::AbstractVDR) = (vdr.flags & 0x04) != 0
+
+# sRecords: 0 = none, 1 = pad sparse, 2 = previous-record sparse
+sparse_type(vdr::AbstractVDR) = Int(vdr.s_records)
+
+# PadValue trails zDimSizes+DimVarys (zVDR) / DimVarys (rVDR), both starting at vdr.pos
+_pad_value_pos(vdr::VDR) = vdr.pos + 8 * Int(vdr.num_dims)
+_pad_value_pos(vdr::rVDR) = vdr.pos + 4 * Int(vdr.gdr.r_num_dims)
+
+# Returned in file encoding: virtual records are filled before the final byte swap
+# in `readblock!`, which fixes pad and physical data together.
+function pad_value(vdr::AbstractVDR, ::Type{T}, needs_byte_swap) where {T}
+    if has_pad_value(vdr)
+        buf = vdr.buffer
+        return GC.@preserve buf unsafe_load(convert(Ptr{T}, pointer(buf, _pad_value_pos(vdr))))
+    end
+    return _file_encode(default_pad(T), needs_byte_swap)
+end
+
+_file_encode(x, needs_byte_swap) = needs_byte_swap ? hton(x) : x
+_file_encode(x::StaticString, needs_byte_swap) = x # excluded from byte swap
+
+# NASA CDF library default pad values (≠ ISTP FILLVAL)
+default_pad(::Type{Int8}) = Int8(-127)
+default_pad(::Type{Int16}) = Int16(-32767)
+default_pad(::Type{Int32}) = Int32(-2147483647)
+default_pad(::Type{Int64}) = Int64(-9223372036854775807)
+default_pad(::Type{UInt8}) = UInt8(254)
+default_pad(::Type{UInt16}) = UInt16(65534)
+default_pad(::Type{UInt32}) = UInt32(4294967294)
+default_pad(::Type{Float32}) = Float32(-1.0e30)
+default_pad(::Type{Float64}) = -1.0e30
+default_pad(::Type{Epoch}) = Epoch(-1.0e30)
+default_pad(::Type{Epoch16}) = Epoch16(-1.0e30, -1.0e30)
+default_pad(::Type{TT2000}) = TT2000(Int64(-9223372036854775807))
+default_pad(::Type{StaticString{N, UInt8}}) where {N} = StaticString{N, UInt8}(ntuple(_ -> UInt8(' '), Val(N)))
diff --git a/test/make_sparse_cdf.py b/test/make_sparse_cdf.py
new file mode 100644
index 0000000..9cd85ea
--- /dev/null
+++ b/test/make_sparse_cdf.py
@@ -0,0 +1,45 @@
+# Generate data/a_sparse_cdf.cdf: sparse-record fixtures (pad & prev semantics).
+# Run: uv run --with cdflib python test/make_sparse_cdf.py
+from pathlib import Path
+
+import numpy as np
+from cdflib import cdfwrite
+
+out = Path(__file__).parent.parent / "data" / "a_sparse_cdf.cdf"
+out.unlink(missing_ok=True)
+f = cdfwrite.CDF(out)
+
+# physical records 0-2, 6-7, 10 (0-based); virtual gaps at 3-5, 8-9
+phys = np.array([0, 1, 2, 6, 7, 10])
+data = np.array([10.0, 11.0, 12.0, 16.0, 17.0, 20.0])
+
+base = {
+    "Data_Type": 45,  # CDF_DOUBLE
+    "Num_Elements": 1,
+    "Rec_Vary": True,
+    "Dim_Sizes": [],
+    "Compress": 0,
+    "Block_Factor": 3,  # force multiple VVR blocks
+}
+f.write_var({**base, "Variable": "pad_default", "Sparse": "pad_sparse"}, var_data=[phys, data])
+f.write_var(
+    {**base, "Variable": "pad_explicit", "Sparse": "pad_sparse", "Pad": np.array([-99.0])},
+    var_data=[phys, data],
+)
+f.write_var({**base, "Variable": "prev", "Sparse": "prev_sparse"}, var_data=[phys, data])
+
+data2d = np.arange(12.0).reshape(6, 2)
+f.write_var(
+    {**base, "Variable": "pad2d", "Dim_Sizes": [2], "Sparse": "pad_sparse", "Pad": np.array([-99.0])},
+    var_data=[phys, data2d],
+)
+f.close()
+
+# sanity-check with the cdflib reader
+import cdflib
+
+r = cdflib.CDF(out)
+print("pad_default:", r.varget("pad_default"))
+print("pad_explicit:", r.varget("pad_explicit"))
+print("prev:", r.varget("prev"))
+print("pad2d:", r.varget("pad2d").tolist())
diff --git a/test/runtests.jl b/test/runtests.jl
index 097ba19..efa6cad 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -8,6 +8,7 @@ include("comprehensive_test.jl")
 include("cdf2_test.jl")
 include("CommonDataModelExt_test.jl")
 include("decompress_test.jl")
+include("sparse_test.jl")
 @testset "StaticString" begin
     include("staticstring.jl")
 end
diff --git a/test/sparse_test.jl b/test/sparse_test.jl
new file mode 100644
index 0000000..fcf86a8
--- /dev/null
+++ b/test/sparse_test.jl
@@ -0,0 +1,32 @@
+# Fixture from test/make_sparse_cdf.py (cdflib writer): physical records 1-3, 7-8, 11
+# (1-based), virtual gaps at 4-6 and 9-10.
+@testset "Sparse records" begin
+    ds = CDFDataset(data_path("a_sparse_cdf.cdf"))
+
+    phys = [10.0, 11.0, 12.0, 16.0, 17.0, 20.0]
+
+    # pad sparse, NASA default pad for CDF_DOUBLE
+    d = -1.0e30
+    @test ds["pad_default"][:] == [10.0, 11.0, 12.0, d, d, d, 16.0, 17.0, d, d, 20.0]
+
+    # pad sparse with explicit VDR pad value
+    p = ds["pad_explicit"]
+    @test p[:] == [10.0, 11.0, 12.0, -99.0, -99.0, -99.0, 16.0, 17.0, -99.0, -99.0, 20.0]
+    @test p[3:5] == [12.0, -99.0, -99.0] # partial range straddling physical/virtual
+    @test p[4:4] == [-99.0] # purely virtual range
+    @test p[11:11] == [20.0]
+
+    # previous sparse: repeat last record of preceding block; pad before first block
+    @test ds["prev"][:] == [10.0, 11.0, 12.0, 12.0, 12.0, 12.0, 16.0, 17.0, 17.0, 17.0, 20.0]
+
+    # 2D: whole virtual record takes the pad value (cdflib's reader only pads the
+    # first element — spec says the full record)
+    m = ds["pad2d"][:, :]
+    @test size(m) == (2, 11)
+    @test m[:, 1] == [0.0, 1.0]
+    @test m[:, 4] == [-99.0, -99.0]
+    @test m[:, 11] == [10.0, 11.0]
+    @test ds["pad2d"][1:1, 5:7] == [-99.0 -99.0 6.0]
+
+    @test read(ds, "prev", Vector{Float64}) == ds["prev"][:]
+end