diff --git a/src/Parser.jl b/src/Parser.jl
index d066672..b7556bb 100644
--- a/src/Parser.jl
+++ b/src/Parser.jl
@@ -3,6 +3,8 @@ module Parser  # JSON
 using Mmap
 using ..Common
 
+include("pushvector.jl")
+
 """
 Like `isspace`, but work on bytes and includes only the four whitespace
 characters defined by the JSON standard: space, tab, line feed, and carriage
@@ -31,9 +33,9 @@ mutable struct StreamingParserState{T <: IO} <: ParserState
     io::T
     cur::UInt8
     used::Bool
-    utf8array::Vector{UInt8}
+    utf8array::PushVector{UInt8, Vector{UInt8}}
 end
-StreamingParserState(io::IO) = StreamingParserState(io, 0x00, true, UInt8[])
+StreamingParserState(io::IO) = StreamingParserState(io, 0x00, true, PushVector{UInt8}())
 
 struct ParserContext{DictType, IntType} end
 
diff --git a/src/pushvector.jl b/src/pushvector.jl
new file mode 100644
index 0000000..01399f1
--- /dev/null
+++ b/src/pushvector.jl
@@ -0,0 +1,33 @@
+# This is a vector wrapper that we use as a workaround for `push!`
+# being slow (it always calls into the runtime even if the underlying buffer,
+# has enough space). Here we keep track of the length using an extra field
+mutable struct PushVector{T, A<:AbstractVector{T}} <: AbstractVector{T}
+    v::A
+    l::Int
+end
+
+# Default length of 20 should be enough to never need to grow in most cases
+PushVector{T}() where {T} = PushVector(Vector{T}(undef, 20), 0)
+
+Base.unsafe_convert(::Type{Ptr{UInt8}}, v::PushVector) = pointer(v.v)
+Base.length(v::PushVector) = v.l
+Base.size(v::PushVector) = (v.l,)
+@inline function Base.getindex(v::PushVector, i)
+    @boundscheck checkbounds(v, i)
+    @inbounds v.v[i]
+end
+
+function Base.push!(v::PushVector, i)
+    v.l += 1
+    if v.l > length(v.v)
+        resize!(v.v, v.l * 2)
+    end
+    v.v[v.l] = i
+    return v
+end
+
+function Base.resize!(v::PushVector, l::Integer)
+    # Only support shrinking for now, since that is all we need
+    @assert l <= v.l
+    v.l = l
+end