diff --git a/experimental/iterators/include/iterators/Conversion/Passes.td b/experimental/iterators/include/iterators/Conversion/Passes.td
index 426acfe4c8b6..cdc4f60d1f33 100644
--- a/experimental/iterators/include/iterators/Conversion/Passes.td
+++ b/experimental/iterators/include/iterators/Conversion/Passes.td
@@ -82,6 +82,10 @@ def ConvertStatesToLLVM : Pass<"convert-states-to-llvm", "ModuleOp"> {
 // TabularToLLVM
 //===----------------------------------------------------------------------===//
 
+// TODO(ingomueller): This pass should probably not exist. Instead, tabular
+//                    views should be lowered into its constituant components,
+//                    which could be tensors, in order to compose better with
+//                    other dialects and passes.
 def ConvertTabularToLLVM : Pass<"convert-tabular-to-llvm", "ModuleOp"> {
   let summary = "Convert the tabular dialect to the LLVM dialect";
   let description = [{
diff --git a/experimental/iterators/include/iterators/Dialect/Iterators/IR/IteratorsOps.td b/experimental/iterators/include/iterators/Dialect/Iterators/IR/IteratorsOps.td
index 3cce9e528483..a03413386650 100644
--- a/experimental/iterators/include/iterators/Dialect/Iterators/IR/IteratorsOps.td
+++ b/experimental/iterators/include/iterators/Dialect/Iterators/IR/IteratorsOps.td
@@ -328,7 +328,12 @@ def Iterators_TabularViewToStreamOp : Iterators_Op<"tabular_view_to_stream",
                     "          .cast<LLVM::LLVMStructType>().getBody())">,
      DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>]> {
   // TODO(ingomueller): Get rid of the LLVM dependency after designing a row
-  //                    type or similar.
+  //                    type or similar. Apart from tying the lowering to LLVM,
+  //                    which probably limits the applicability of higher-level
+  //                    passes on the components of a tabular view, the
+  //                    LLVM-based design also inherits the limitations of the
+  //                    LLVM type system, such as the missing support for signed
+  //                    and unsigned integer types.
   let summary = "Extracts the tuples from a tabular view as LLVM structs";
   let description = [{
     Produces a stream of LLVM structs from a given `tabular_view` (i.e., "scans"
diff --git a/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularDialect.td b/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularDialect.td
index 8e9d3e7c1572..6abb804c537f 100644
--- a/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularDialect.td
+++ b/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularDialect.td
@@ -15,6 +15,18 @@ include "mlir/IR/OpBase.td"
 // Dialect definition
 //===----------------------------------------------------------------------===//
 
+// TODO(ingomueller): This dialect should probably not be lowered to LLVM
+//                    directly. I think it would be a better design to treat
+//                    tabular views (potentially with a different name) as a
+//                    bundle of several tensors, which we can access (and maybe
+//                    update) one at a time. It is then up to the consumer to
+//                    extract rows or batches of rows from the tensors, for
+//                    which they can use the usual means. The default lowering
+//                    would then simple decompose the tabular view into its
+//                    constituant tensors, which would not affect the accesses
+//                    to rows or row batches. After that decomposition, normal
+//                    bufferization can run without knowing the tabular
+//                    provenance of the tensors.
 def Tabular_Dialect : Dialect {
   let name = "tabular";
   let cppNamespace = "::mlir::iterators";
diff --git a/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularOps.td b/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularOps.td
index 2849c1da2885..f48d5a066099 100644
--- a/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularOps.td
+++ b/experimental/iterators/include/iterators/Dialect/Tabular/IR/TabularOps.td
@@ -34,6 +34,10 @@ class MemRefTypeWithIdentityLayoutAndRankOf<list<Type> allowedTypes,
 
 def Tabular_ViewAsTabularOp : Tabular_Op<"view_as_tabular",
     [DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>]> {
+  // TODO(ingomueller): We should consider allowing tensors as operands, or even
+  //                    make that the only possibility. This would allow to keep
+  //                    the highest level of abstraction, in particular, if we
+  //                    do not lower to LLVM directly.
   let summary = "Creates a `tabular_view` from the given memrefs";
   let description = [{
     Converts a variadic number of memrefs of rank 1 into the columns of a
@@ -70,4 +74,8 @@ def Tabular_ViewAsTabularOp : Tabular_Op<"view_as_tabular",
   }];
 }
 
+// TODO(ingomueller): It may make sense to add accessors for individual columns,
+//                    i.e., an ExtractColumnOp or similar, which would return a
+//                    single column of the tabular view, possibly a tensor.
+
 #endif // TABULAR_DIALECT_TABULAR_IR_TABULAROPS